00 mi piace00 non mi piace

53 visualizzazioni593 pagineThanks

Mar 18, 2017

© © All Rights Reserved

PDF, TXT o leggi online da Scribd

Thanks

© All Rights Reserved

53 visualizzazioni

00 mi piace00 non mi piace

Thanks

© All Rights Reserved

Sei sulla pagina 1di 593

Second Edition

ii Contents

Contents iii

Second Edition

George G. Roussas

Intercollege Division of Statistics

University of California

Davis, California

ACADEMIC PRESS

San Diego London Boston

New York Sydney Tokyo Toronto

iv Contents

No part of this publication may be reproduced or transmitted in any form or by

any means, electronic or mechanical, including photocopy, recording, or any

information storage and retrieval system, without permission in writing from the

publisher.

ACADEMIC PRESS

525 B Street, Suite 1900, San Diego, CA 92101-4495, USA

1300 Boylston Street, Chestnut Hill, MA 02167, USA

http://www.apnet.com

2428 Oval Road, London NW1 7DX, UK

http://www.hbuk.co.uk/ap/

Roussas, George G.

A course in mathematical statistics / George G. Roussas.2nd ed.

p. cm.

Rev. ed. of: A first course in mathematical statistics. 1973.

Includes index.

ISBN 0-12-599315-3

1. Mathematical statistics. I. Roussas, George G. First course in mathematical

statistics. II. Title.

QA276.R687 1997 96-42115

519.5dc20 CIP

96 97 98 99 00 EB 9 8 7 6 5 4 3 2 1

Contents v

vi Contents

Contents vii

Contents

Exercises 5

1.2* Fields and -Fields 8

2.1 Probability Functions and Some Basic Properties and Results 14

Exercises 20

2.2 Conditional Probability 21

Exercises 25

2.3 Independence 27

Exercises 33

2.4 Combinatorial Results 34

Exercises 40

2.5* Product Probability Spaces 45

Exercises 47

vii

viii Contents

Exercises 52

3.2 Discrete Random Variables (and Random Vectors) 55

Exercises 61

3.3 Continuous Random Variables (and Random Vectors) 65

Exercises 76

3.4 The Poisson Distribution as an Approximation to the Binomial

Distribution and the Binomial Distribution as an Approximation to the

Hypergeometric Distribution 79

Exercises 82

3.5* Random Variables as Measurable Functions and Related Results 82

Exercises 84

Relationship 85

VectorBasic Properties of the d.f. of a Random Variable 85

Exercises 89

4.2 The d.f. of a Random Vector and Its PropertiesMarginal and

Conditional d.f.s and p.d.f.s 91

Exercises 97

4.3 Quantiles and Modes of a Distribution 99

Exercises 102

4.4* Justification of Statements 1 and 2 102

Exercises 105

Probability Inequalities 106

Exercises 111

5.2 Expectations and Variances of Some R.V.s 114

Exercises 119

5.3 Conditional Moments of Random Variables 122

Exercises 124

Contents ix

Inequalities 125

Exercises 128

5.5 Covariance, Correlation Coefficient and Its Interpretation 129

Exercises 133

5.6* Justification of Relation (2) in Chapter 2 134

and Related Theorems 138

6.1 Preliminaries 138

6.2 Definitions and Basic TheoremsThe One-Dimensional Case 140

Exercises 145

6.3 The Characteristic Functions of Some Random Variables 146

Exercises 149

6.4 Definitions and Basic TheoremsThe Multidimensional Case 150

Exercises 152

6.5 The Moment Generating Function and Factorial Moment Generating

Function of a Random Variable 153

Exercises 160

7.1 Stochastic Independence: Criteria of Independence 164

Exercises 168

7.2 Proof of Lemma 2 and Related Results 170

Exercises 172

7.3 Some Consequences of Independence 173

Exercises 176

7.4* Independence of Classes of Events and Related Results 177

Exercise 179

8.1 Some Modes of Convergence 180

Exercises 182

8.2 Relationships Among the Various Modes of Convergence 182

Exercises 187

8.3 The Central Limit Theorem 187

Exercises 194

8.4 Laws of Large Numbers 196

Exercises 198

x Contents

Exercises 206

8.6* Plyas Lemma and Alternative Proof of the WLLN 206

Exercises 211

Vectors 212

9.1 The Univariate Case 212

Exercises 218

9.2 The Multivariate Case 219

Exercises 233

9.3 Linear Transformations of Random Vectors 235

Exercises 240

9.4 The Probability Integral Transform 242

Exercise 244

10.1 Order Statistics and Related Distributions 245

Exercises 252

10.2 Further Distribution Theory: Probability of Coverage of a Population

Quantile 256

Exercise 258

11.1 Sufficiency: Definition and Some Basic Results 260

Exercises 269

11.2 Completeness 271

Exercises 273

11.3 UnbiasednessUniqueness 274

Exercises 276

11.4 The Exponential Family of p.d.f.s: One-Dimensional Parameter

Case 276

Exercises 280

11.5 Some Multiparameter Generalizations 281

Exercises 282

12.1 Introduction 284

Exercise 284

Contents xi

Variance 285

Exercises 286

12.3 The Case of Availability of Complete Sufficient Statistics 287

Exercises 292

12.4 The Case Where Complete Sufficient Statistics Are Not Available or

May Not Exist: Cramr-Rao Inequality 293

Exercises 301

12.5 Criteria for Selecting an Estimator: The Maximum Likelihood

Principle 302

Exercises 308

12.6 Criteria for Selecting an Estimator: The Decision-

Theoretic Approach 309

12.7 Finding Bayes Estimators 312

Exercises 317

12.8 Finding Minimax Estimators 318

Exercise 320

12.9 Other Methods of Estimation 320

Exercises 322

12.10 Asymptotically Optimal Properties of Estimators 322

Exercise 325

12.11 Closing Remarks 325

Exercises 326

13.1 General Concepts of the Neyman-Pearson Testing Hypotheses

Theory 327

Exercise 329

13.2 Testing a Simple Hypothesis Against a Simple Alternative 329

Exercises 336

13.3 UMP Tests for Testing Certain Composite Hypotheses 337

Exercises 347

13.4 UMPU Tests for Testing Certain Composite Hypotheses 349

Exercises 353

13.5 Testing the Parameters of a Normal Distribution 353

Exercises 356

13.6 Comparing the Parameters of Two Normal Distributions 357

Exercises 360

13.7 Likelihood Ratio Tests 361

Exercises 369

13.8 Applications of LR Tests: Contingency Tables, Goodness-of-Fit

Tests 370

Exercises 373

xii Contents

14.1 Some Basic Theorems of Sequential Sampling 382

Exercises 388

14.2 Sequential Probability Ratio Test 388

Exercise 392

14.3 Optimality of the SPRT-Expected Sample Size 393

Exercises 394

14.4 Some Examples 394

15.1 Confidence Intervals 397

Exercise 398

15.2 Some Examples 398

Exercises 404

15.3 Confidence Intervals in the Presence of Nuisance Parameters 407

Exercise 410

15.4 Confidence RegionsApproximate Confidence Intervals 410

Exercises 412

15.5 Tolerance Intervals 413

16.1 Introduction of the Model 416

16.2 Least Square EstimatorsNormal Equations 418

16.3 Canonical Reduction of the Linear ModelEstimation of 2 424

Exercises 428

16.4 Testing Hypotheses About = E(Y) 429

Exercises 433

16.5 Derivation of the Distribution of the F Statistic 433

Exercises 436

17.1 One-way Layout (or One-way Classification) with the Same Number of

Observations Per Cell 440

Exercise 446

17.2 Two-way Layout (Classification) with One Observation Per Cell 446

Exercises 451

Contents xiii

17.3 2) Observations

Two-way Layout (Classification) with K (

Per Cell 452

Exercises 457

17.4 A Multicomparison method 458

Exercises 462

18.1 Introduction 463

Exercises 466

18.2 Some Properties of Multivariate Normal Distributions 467

Exercise 469

18.3 Estimation of and / and a Test of Independence 469

Exercises 475

19.1 Introduction 476

19.2 Some Theorems on Quadratic Forms 477

Exercises 483

20.1 Nonparametric Estimation 485

20.2 Nonparametric Estimation of a p.d.f. 487

Exercise 490

20.3 Some Nonparametric Tests 490

20.4 More About Nonparametric Tests: Rank Tests 493

Exercises 496

20.5 Sign Test 496

20.6 Relative Asymptotic Efficiency of Tests 498

I.1 Basic Definitions in Vector Spaces 499

I.2 Some Theorems on Vector Spaces 501

I.3 Basic Definitions About Matrices 502

I.4 Some Theorems About Matrices and Quadratic Forms 504

II.1 Noncentral t-Distribution 508

xiv Contents

II.3 Noncentral F-Distribution 509

2 The Cumulative Poisson Distribution 520

3 The Normal Distribution 523

4 Critical Values for Students t-Distribution 526

5 Critical Values for the Chi-Square Distribution 529

6 Critical Values for the F-Distribution 532

7 Table of Selected Discrete and Continuous Distributions and Some of

Their Characteristics 542

Index 561

Contents xv

This is the second edition of a book published for the first time in 1973 by

Addison-Wesley Publishing Company, Inc., under the title A First Course in

Mathematical Statistics. The first edition has been out of print for a number of

years now, although its reprint in Taiwan is still available. That issue, however,

is meant for circulation only in Taiwan.

The first issue of the book was very well received from an academic

viewpoint. I have had the pleasure of hearing colleagues telling me that the

book filled an existing gap between a plethora of textbooks of lower math-

ematical level and others of considerably higher level. A substantial number of

colleagues, holding senior academic appointments in North America and else-

where, have acknowledged to me that they made their entrance into the

wonderful world of probability and statistics through my book. I have also

heard of the book as being in a class of its own, and also as forming a collectors

item, after it went out of print. Finally, throughout the years, I have received

numerous inquiries as to the possibility of having the book reprinted. It is in

response to these comments and inquiries that I have decided to prepare a

second edition of the book.

This second edition preserves the unique character of the first issue of the

book, whereas some adjustments are affected. The changes in this issue consist

in correcting some rather minor factual errors and a considerable number of

misprints, either kindly brought to my attention by users of the book or

located by my students and myself. Also, the reissuing of the book has pro-

vided me with an excellent opportunity to incorporate certain rearrangements

of the material.

One change occurring throughout the book is the grouping of exercises of

each chapter in clusters added at the end of sections. Associating exercises

with material discussed in sections clearly makes their assignment easier. In

the process of doing this, a handful of exercises were omitted, as being too

complicated for the level of the book, and a few new ones were inserted. In

xv

xvi Contents

Preface to the Second Edition

Chapters 1 through 8, some of the materials were pulled out to form separate

sections. These sections have also been marked by an asterisk (*) to indicate

the fact that their omission does not jeopardize the flow of presentation and

understanding of the remaining material.

Specifically, in Chapter 1, the concepts of a field and of a -field, and basic

results on them, have been grouped together in Section 1.2*. They are still

readily available for those who wish to employ them to add elegance and rigor

in the discussion, but their inclusion is not indispensable. In Chapter 2, the

number of sections has been doubled from three to six. This was done by

discussing independence and product probability spaces in separate sections.

Also, the solution of the problem of the probability of matching is isolated in a

section by itself. The section on the problem of the probability of matching and

the section on product probability spaces are also marked by an asterisk for the

reason explained above. In Chapter 3, the discussion of random variables as

measurable functions and related results is carried out in a separate section,

Section 3.5*. In Chapter 4, two new sections have been created by discussing

separately marginal and conditional distribution functions and probability

density functions, and also by presenting, in Section 4.4*, the proofs of two

statements, Statements 1 and 2, formulated in Section 4.1; this last section is

also marked by an asterisk. In Chapter 5, the discussion of covariance and

correlation coefficient is carried out in a separate section; some additional

material is also presented for the purpose of further clarifying the interpreta-

tion of correlation coefficient. Also, the justification of relation (2) in Chapter 2

is done in a section by itself, Section 5.6*. In Chapter 6, the number of sections

has been expanded from three to five by discussing in separate sections charac-

teristic functions for the one-dimensional and the multidimensional case, and

also by isolating in a section by itself definitions and results on moment-

generating functions and factorial moment generating functions. In Chapter 7,

the number of sections has been doubled from two to four by presenting the

proof of Lemma 2, stated in Section 7.1, and related results in a separate

section; also, by grouping together in a section marked by an asterisk defini-

tions and results on independence. Finally, in Chapter 8, a new theorem,

Theorem 10, especially useful in estimation, has been added in Section 8.5.

Furthermore, the proof of Plyas lemma and an alternative proof of the Weak

Law of Large Numbers, based on truncation, are carried out in a separate

section, Section 8.6*, thus increasing the number of sections from five to six.

In the remaining chapters, no changes were deemed necessary, except that

in Chapter 13, the proof of Theorem 2 in Section 13.3 has been facilitated by

the formulation and proof in the same section of two lemmas, Lemma 1 and

Lemma 2. Also, in Chapter 14, the proof of Theorem 1 in Section 14.1 has been

somewhat simplified by the formulation and proof of Lemma 1 in the same

section.

Finally, a table of some commonly met distributions, along with their

means, variances and other characteristics, has been added. The value of such

a table for reference purposes is obvious, and needs no elaboration.

Preface to the Second

Contents

Edition xvii

This book contains enough material for a year course in probability and

statistics at the advanced undergraduate level, or for first-year graduate stu-

dents not having been exposed before to a serious course on the subject

matter. Some of the material can actually be omitted without disrupting the

continuity of presentation. This includes the sections marked by asterisks,

perhaps, Sections 13.413.6 in Chapter 13, and all of Chapter 14. The instruc-

tor can also be selective regarding Chapters 11 and 18. As for Chapter 19, it

has been included in the book for completeness only.

The book can also be used independently for a one-semester (or even one

quarter) course in probability alone. In such a case, one would strive to cover

the material in Chapters 1 through 10 with the exclusion, perhaps, of the

sections marked by an asterisk. One may also be selective in covering the

material in Chapter 9.

In either case, presentation of results involving characteristic functions

may be perfunctory only, with emphasis placed on moment-generating func-

tions. One should mention, however, why characteristic functions are intro-

duced in the first place, and therefore what one may be missing by not utilizing

this valuable tool.

In closing, it is to be mentioned that this author is fully aware of the fact

that the audience for a book of this level has diminished rather than increased

since the time of its first edition. He is also cognizant of the trend of having

recipes of probability and statistical results parading in textbooks, depriving

the reader of the challenge of thinking and reasoning instead delegating the

thinking to a computer. It is hoped that there is still room for a book of the

nature and scope of the one at hand. Indeed, the trend and practices just

described should make the availability of a textbook such as this one exceed-

ingly useful if not imperative.

G. G. Roussas

Davis, California

May 1996

xviii Contents

undergraduate level, as well as for first-year graduate students in statisticsor

graduate students, in generalwith no prior knowledge of statistics. A typical

three-semester course in calculus and some familiarity with linear algebra

should suffice for the understanding of most of the mathematical aspects of

this book. Some advanced calculusperhaps taken concurrentlywould be

helpful for the complete appreciation of some fine points.

There are basically two streams of textbooks on mathematical statistics

that are currently on the market. One category is the advanced level texts

which demonstrate the statistical theories in their full generality and math-

ematical rigor; for that purpose, they require a high level, mathematical back-

ground of the reader (for example, measure theory, real and complex

analysis). The other category consists of intermediate level texts, where the

concepts are demonstrated in terms of intuitive reasoning, and results are

often stated without proofs or with partial proofs that fail to satisfy an inquisi-

tive mind. Thus, readers with a modest background in mathematics and a

strong motivation to understand statistical concepts are left somewhere in

between. The advanced texts are inaccessible to them, whereas the intermedi-

ate texts deliver much less than they hope to learn in a course of mathematical

statistics. The present book attempts to bridge the gap between the two

categories, so that students without a sophisticated mathematical background

can assimilate a fairly broad spectrum of the theorems and results from math-

ematical statistics. This has been made possible by developing the fundamen-

tals of modern probability theory and the accompanying mathematical ideas at

the beginning of this book so as to prepare the reader for an understanding of

the material presented in the later chapters.

This book consists of two parts, although it is not formally so divided. Part

1 (Chapters 110) deals with probability and distribution theory, whereas Part

xviii

Contents

Preface to the First Edition xix

the concepts of a field and -field, and also the definition of a random variable

as a measurable function, are introduced. This allows us to state and prove

fundamental results in their full generality that would otherwise be presented

vaguely using statements such as it may be shown that . . . , it can be proved

that . . . , etc. This we consider to be one of the distinctive characteristics of

this part. Other important features are as follows: a detailed and systematic

discussion of the most useful distributions along with figures and various

approximations for several of them; the establishment of several moment and

probability inequalities; the systematic employment of characteristic func-

tionsrather than moment generating functionswith all the well-known

advantages of the former over the latter; an extensive chapter on limit theo-

rems, including all common modes of convergence and their relationship; a

complete statement and proof of the Central Limit Theorem (in its classical

form); statements of the Laws of Large Numbers and several proofs of the

Weak Law of Large Numbers, and further useful limit theorems; and also an

extensive chapter on transformations of random variables with numerous

illustrative examples discussed in detail.

The second part of the book opens with an extensive chapter on suffi-

ciency. The concept of sufficiency is usually treated only in conjunction with

estimation and testing hypotheses problems. In our opinion, this does not

do justice to such an important concept as that of sufficiency. Next, the point

estimation problem is taken up and is discussed in great detail and as

large a generality as is allowed by the level of this book. Special attention is

given to estimators derived by the principles of unbiasedness, uniform mini-

mum variance and the maximum likelihood and minimax principles. An abun-

dance of examples is also found in this chapter. The following chapter is

devoted to testing hypotheses problems. Here, along with the examples (most

of them numerical) and the illustrative figures, the reader finds a discussion of

families of probability density functions which have the monotone likelihood

ratio property and, in particular, a discussion of exponential families. These

latter topics are available only in more advanced texts. Other features are

a complete formulation and treatment of the general Linear Hypothesis

and the discussion of the Analysis of Variance as an application of it.

In many textbooks of about the same level of sophistication as the present

book, the above two topics are approached either separately or in the reverse

order from the one used here, which is pedagogically unsound, although

historically logical. Finally, there are special chapters on sequential proce-

dures, confidence regionstolerance intervals, the Multivariate Normal distri-

bution, quadratic forms, and nonparametric inference.

A few of the proofs of theorems and some exercises have been drawn from

recent publications in journals.

For the convenience of the reader, the book also includes an appendix

summarizing all necessary results from vector and matrix algebra.

There are more than 120 examples and applications discussed in detail in

xx Contentsto the First Edition

Preface

the text. Also, there are more than 530 exercises, appearing at the end of the

chapters, which are of both theoretical and practical importance.

The careful selection of the material, the inclusion of a large variety of

topics, the abundance of examples, and the existence of a host of exercises of

both theoretical and applied nature will, we hope, satisfy people of both

theoretical and applied inclinations. All the application-oriented reader has to

do is to skip some fine points of some of the proofs (or some of the proofs

altogether!) when studying the book. On the other hand, the careful handling

of these same fine points should offer some satisfaction to the more math-

ematically inclined readers.

The material of this book has been presented several times to classes

of the composition mentioned earlier; that is, classes consisting of relatively

mathematically immature, eager, and adventurous sophomores, as well as

juniors and seniors, and statistically unsophisticated graduate students. These

classes met three hours a week over the academic year, and most of the

material was covered in the order in which it is presented with the occasional

exception of Chapters 14 and 20, Section 5 of Chapter 5, and Section 3 of

Chapter 9. We feel that there is enough material in this book for a three-

quarter session if the classes meet three or even four hours a week.

At various stages and times during the organization of this book several

students and colleagues helped improve it by their comments. In connection

with this, special thanks are due to G. K. Bhattacharyya. His meticulous

reading of the manuscripts resulted in many comments and suggestions that

helped improve the quality of the text. Also thanks go to B. Lind, K. G.

Mehrotra, A. Agresti, and a host of others, too many to be mentioned here. Of

course, the responsibility in this book lies with this author alone for all omis-

sions and errors which may still be found.

As the teaching of statistics becomes more widespread and its level of

sophistication and mathematical rigor (even among those with limited math-

ematical training but yet wishing to know why and how) more demanding,

we hope that this book will fill a gap and satisfy an existing need.

G. G. R.

Madison, Wisconsin

November 1972

1.1 Some Definitions and Notation 1

Chapter 1

A set S is a (well defined) collection of distinct objects which we denote by s.

The fact that s is a member of S, an element of S, or that it belongs to S is

expressed by writing s S. The negation of the statement is expressed by

writing s S. We say that S is a subset of S, or that S is contained in S, and

write S S, if for every s S, we have s S. S is said to be a proper subset

of S, and we write S S, if S S and there exists s S such that s S. Sets

are denoted by capital letters, while lower case letters are used for elements of

sets.

S

Figure 1.1 S S; in fact, S S, since s2 S,

S'

s1 but s2 S .

s2

diagram (Fig. 1.1). From now on a basic, or universal set, or space (which may

be different from situation to situation), to be denoted by S, will be considered

and all other sets in question will be subsets of S.

1. The complement (with respect to S) of the set A, denoted by Ac, is

defined by Ac = {s S; s A}. (See Fig. 1.2.)

1

2 1 Basic Concepts of Set Theory

S

A

Ac Figure 1.2 Ac is the shaded region.

n

A1 A2 An or U Aj ,

j =1

is defined by

n

j =1

infinite number of sets. Thus for denumerably many sets, one has

j =1

A1 A2

n

A1 A2 An or I Aj ,

j =1

is defined by

n

j =1

infinite number of sets. Thus for denumerably many sets, one has

I A j = {s S ; s A j for all j = 1, 2, }.

j =1

A1 A2

1.1 Some Definitions and Notation 3

{

A1 A2 = s S ; s A1 , s A2 . }

Symmetrically,

{

A2 A1 = s S ; s A2 , s A1 . }

Note that A1 A2 = A1 Ac2, A2 A1 = A2 Ac1, and that, in general, A1 A 2

A2 A1. (See Fig. 1.5.)

S

Figure 1.5 A1 A2 is ////.

A2 A1 is \\\\.

A1 A2

( ) (

A1 A2 = A1 A2 A2 A1 . )

Note that

( ) (

A1 A2 = A1 A2 A1 A2 . )

Pictorially, this is shown in Fig. 1.6. It is worthwhile to observe that

operations (4) and (5) can be expressed in terms of operations (1), (2), and

(3).

A1 A2

A set which contains no elements is called the empty set and is denoted by .

Two sets A1, A2 are said to be disjoint if A1 A2 = . Two sets A1, A2 are said

to be equal, and we write A1 = A2, if both A1 A2 and A2 A1. The sets Aj,

j = 1, 2, . . . are said to be pairwise or mutually disjoint if Ai Aj = for all

i j (Fig. 1.7). In such a case, it is customary to write

n

A1 + A2 , A1 + + An = A j and A1 + A2 + = A j

j =1 j =1

instead of A1 A2, U A j, and U A j, respectively. We will write

n

j =1 j =1

U j A j, j A j, I j A j, where we do not wish to specify the range of j, which

4 1 Basic Concepts of Set Theory

will usually be either the (finite) set {1, 2, . . . , n}, or the (infinite) set

{1, 2, . . .}.

S

Figure 1.7 A1 and A2 are disjoint; that is,

A1 A2 = . Also A1 A2 = A1 + A2 for the

same reason.

A1 A2

1. S c = , c = S, (Ac)c = A.

2. S A = S, A = A, A Ac = S, A A = A.

3. S A = A, A = , A Ac = , A A = A.

The previous statements are all obvious as is the following: A for every

subset A of S. Also

4. A1 (A2 A3) = (A1 A2) A3

A1 (A2 A3) = (A1 A2) A3 } (Associative laws)

5. A1 A2 = A2 A1

A1 A2 = A2 A1 } (Commutative laws)

6. A (j Aj) = j (A Aj)

A (j Aj) = j (A Aj) } (Distributive laws)

The following identity is a useful tool in writing a union of sets as a sum of

disjoint sets.

An identity:

j

There are two more important properties of the operation on sets which

relate complementation to union and intersection. They are known as De

Morgans laws:

c

i ) U Aj = I Aj ,

j j

c

c

ii ) I Aj = U Aj .

j j

c

PROOF OF (i) We wish to establish

1.1 Some Definitions andExercises

Notation 5

We will then, by definition, have verified the desired equality of the two

sets.

a) Let s ( U jAj)c. Then s U jAj, hence s Aj for any j. Thus s Acj for every

j and therefore s I jAcj.

b) Let s I jAcj. Then s Acj for every j and hence s Aj for any j. Then

s U jAj and therefore s ( U jAj)c.

The proof of (ii) is quite similar.

DEFINITION 1 The sequence {An}, n = 1, 2, . . . , is said to be a monotone sequence of sets if:

ii) A1 A2 A3 (that is, An is increasing, to be denoted by An), or

ii) A1 ! A2 ! A3 ! (that is, An is decreasing, to be denoted by An).

The limit of a monotone sequence is defined as follows:

ii) If An, then lim An = U An , and

n

n=1

ii) If An, then lim An = I An .

n

n=1

A = lim inf An = U I Aj,

n

n=1 j= n

and

A = lim sup An = I U Aj .

n n=1 j= n

The sets A and A are called the inferior limit and superior limit,

.

respectively, of the sequence {An}. The sequence {An} has a limit if A = A

Exercises

1.1.1 Let Aj, j = 1, 2, 3 be arbitrary subsets of S. Determine whether each of

the following statements is correct or incorrect.

iii) (A1 A2) A2 = A2;

iii) (A1 A2) A1 = A2;

iii) (A1 A2) (A1 A2) = ;

iv) (A1 A2) (A2 A3) (A3 A1) = (A1 A2) (A2 A3) (A3 A1).

6 1 Basic Concepts of Set Theory

prime denotes transpose, and define the subsets Aj, j = 1, . . . , 7 of S as follows:

( )

A1 = x, y S ; x = y;

( ) S ; x = y;

A2 = x, y

( )

A3 = x, y S ; x 2 = y 2 ;

( ) S ; x

A4 = x, y

2

y 2 ;

( )

A5 = x, y S ; x 2 + y 2 4 ;

( )

A6 = x, y S ; x y2 ;

( )

A7 = x, y S ; x 2 y.

1.1.3 Refer to Exercise 1.1.2 and show that:

7 7

(

iii) A1 U A j = U A1 A j ;

j =2 j =2

)

7 7

(

iii) A1 I A j = I A1 A j ;

j =2 j =2

)

c

7 7

iii) U A j = I Acj ;

j =1 j =1

c

7 7

iv) I A j = U Acj

j =1 j =1

by listing the members of each one of the eight sets appearing on either side of

each one of the relations (i)(iv).

Then show that A C; that is, the subset relationship is transitive. Verify it by

taking A = A1, B = A3 and C = A4, where A1,A3 and A4 are defined in Exercise

1.1.2.

1.1.6 In terms of the acts A1, A2, A3, and perhaps their complements,

express each one of the following acts:

iii) C = {s S; s belongs to all of A1, A2, A3 };

1.1 Some Definitions andExercises

Notation 7

iv) E = {s S; s belongs to at most 2 of A1, A2, A3};

iv) F = {s S; s belongs to at least 1 of A1, A2, A3 }.

1.1.8 Give a detailed proof of the second identity in De Morgans laws; that

is, show that

c

I Aj = U Aj .

c

j j

iii) A = {s S; s belongs to all but finitely many As};

iii) A = {s S; s belongs to infinitely many As};

;

iii) A A

iv) If {An} is a monotone sequence, then A = A = lim An .

n

follows:

1

( ) 1 2

An = x, y ! 2 ; 3 + x < 6 , 0 y 2 2 ,

n n n

1

( )

Bn = x, y ! 2 ; x 2 + y 2 3 .

n

1.1.11 Let S = ! and define the subsets An, Bn, n = 1, 2, . . . of S as

follows:

1 1 3

An = x ! ; 5 + < x < 20 , Bn = x ! ; 0 < x < 7 + .

n n n

An = A and lim B = B exist (by

n n

Exercise 1.1.9(iv)). Also identify the sets A and B.

1.1.12 Let A and B be subsets of S and for n = 1, 2, . . . , define the sets An as

follows: A2n1 = A, A2n = B. Then show that

n n

8 1 Basic Concepts of Set Theory

In this section, we introduce the concepts of a field and of a -field, present a

number of examples, and derive some basic results.

DEFINITION 2 A class (set) of subsets of S is said to be a field, and is denoted by F, if

(F1) F is a non-empty class.

(F2) A F implies that Ac F (that is, F is closed under

complementation).

(F3) A1, A2 F implies that A1 A2 F (that is, F is closed under

pairwise unions).

1. S, F.

2. If Aj F, j = 1, 2, . . . , n, then Unj =1 A j F, Inj =1 A j F for any finite n.

(That is, F is closed under finite unions and intersections. Notice, how-

ever, that Aj F, j = 1, 2, . . . need not imply that their union or intersection is

in F; for a counterexample, see consequence 2 on page 10.)

PROOF OF (1) AND (2) (1) (F1) implies that there exists A F and (F2)

implies that Ac F. By (F3), A Ac = S F. By (F2), S c = F.

(2) The proof will be by induction on n and by one of the De Morgans

laws. By (F3), if A1, A2 F, then A1 A2 F; hence the statement for unions

is true for n = 2. (It is trivially true for n = 1.) Now assume the statement for

unions is true for n = k 1; that is, if

k1

A1 , A2 , , Ak1 F , then U Aj F .

j =1

k k1

U j U A j Ak .

A =

j =1 j =1

k1 k

U A j Ak = U A j F

j =1 j =1

and by induction, the statement for unions is true for any finite n. By observing

that

c

n n c

=

I j U A j ,

A

j =1 j =1

* The reader is reminded that sections marked by an asterisk may be omitted without jeo-

* pardizing the understanding of the remaining material.

1.1 Some1.2*

Definitions

Fields and -Fields

and Notation 9

we see that (F2) and the above statement for unions imply that if A1, . . . , An

F, then Inj =1 A j F for any finite n.

1. C1 = {, S} is a field (trivial field).

2. C2 = {all subsets of S} is a field (discrete field).

3. C3 = {, S, A, Ac}, for some A S, is a field.

4. Let S be infinite (countably so or not) and let C4 be the class of subsets

of S which are finite, or whose complements are finite; that is, C4 = {A S; A

or Ac is finite}.

As an example, we shall verify that C4 is a field.

PROOF THAT C4 IS A FIELD

i) Since S c = is finite, S C4, so that C4 is non-empty.

ii) Suppose that A C4. Then A or Ac is finite. If A is finite, then (Ac)c = A is

finite and hence Ac C4 also. If Ac is finite, then Ac C4.

iii) Suppose that A1, A2 C4. Then A1 or Ac1 is finite and A2 or Ac2 is finite.

A2 C4.

b) Suppose that A1c, A2 are finite. Then (A1 A2)c = A1c Ac2 is finite

since A1c is. Hence A1 A2 C4.

The other two possibilities follow just as in (b). Hence (F1), (F2), (F3) are

satisfied.

We now formulate and prove the following theorems about fields.

THEOREM 1 Let I be any non-empty index set (finite, or countably infinite, or uncoun-

table), and let Fj, j I be fields of subsets of S. Define F by F = I j I F j =

{A; A Fj for all j I}. Then F is a field.

PROOF

i) S, Fj for every j I, so that S, F and hence F is non-empty.

ii) If A F, then A Fj for every j I. Thus Ac Fj for every j I, so that

Ac F.

iii) If A1, A2 F, then A1, A2 Fj for every j I. Then A1 A2 Fj for every

j I, and hence A1 A2 F.

THEOREM 2 Let C be an arbitrary class of subsets of S. Then there is a unique minimal field

F containing C. (We say that F is generated by C and write F = F(C).)

PROOF Clearly, C is contained in the discrete field. Next, let {Fj, j I} be the

class of all fields containing C and define F(C) by

( )

F C = I F j.

j I

10 1 Basic Concepts of Set Theory

field, since it is the intersection of all fields containing C, and is unique. Hence

F = F(C).

DEFINITION 3 A class of subsets of S is said to be a -field, and is denoted by A, if it is a field

and furthermore (F3) is replaced by (A3): If Aj A, j = 1, 2, . . . , then U j = 1 A j

A (that is, A is closed under denumerable unions).

1. If Aj A, j = 1, 2, . . . , then I j = 1 A j A (that is, A is closed under

denumerable intersections).

2. By definition, a -field is a field, but the converse is not true. In fact, in

Example 4 on page 9, take S = (, ), and define Aj = {all integers in [j, j]},

j = 1, 2, . . . . Then U j = 1 A j is the set A, say, of all integers. Thus A is infinite and

furthermore so is Ac. Hence A F, whereas Aj F for all j.

1. C1 = {, S} is a -field (trivial -field).

2. C2 = {all subsets of S} is a -field (discrete -field).

3. C3 = {, S, A, Ac} for some A S is a -field.

4. Take S to be uncountable and define C4 as follows:

C4 = {all subsets of S which are countable or whose complements are

countable}.

As an example, we prove that C4 is a -field.

PROOF

i) Sc = is countable, so C4 is non-empty.

ii) If A C4, then A or Ac is countable. If A is countable, then (Ac)c = A is

countable, so that Ac C4. If Ac is countable, by definition Ac C4.

iii) The proof of this statement requires knowledge of the fact that a count-

able union of countable sets is countable. (For proof of this fact see

page 36 in Tom M. Apostols book Mathematical Analysis, published

by Addison-Wesley, 1957.) Let Aj, j = 1, 2, . . . A. Then either each

Aj is countable, or there exists some Aj for which Aj is not countable but

Acj is. In the first case, we invoke the previously mentioned theorem

on the countable union of countable sets. In the second case, we note

that

c

U j A = I Acj ,

j = 1 j = 1

which is countable, since it is the intersection of sets, one of which is

countable.

1.1 Some1.2*

Definitions

Fields and -Fields

and Notation 11

THEOREM 3 Let I be as in Theorem 1, and let Aj, j I, be -fields. Define A by A =

I jI Aj = {A; A Aj for all j I}. Then A is a -field.

PROOF

i) S, Aj for every j I and hence they belong in A.

ii) If A A, then A Aj for every j I, so that Ac Aj for every j I. Thus

Ac A.

iii) If A1, A2, . . . , A, then A1, A2, . . . Aj for every j I and hence U j = 1 A j

A j; for every j I; thus U j = 1 A j A .

-field A containing C. (We say that A is the -field generated by C and write

A = (C).)

( ) {

C = I all -fields containing C . }

By Theorem 3, (C ) is a -field which obviously contains C. Uniqueness

and minimality again follow from the definition of (C ). Hence A =

(C).

= {C; C = B A for some B A} is a -field, where complements of sets are

formed with respect to A, which now plays the role of the entire space. This is

easily seen to be true by the distributive property of intersection over union

(see also Exercise 1.2.5).

In all that follows, if S is countable (that is, finite or denumerably in-

finite), we will always take A to be the discrete -field. However, if S is

uncountable, then for certain technical reasons, we take the -field to be

smaller than the discrete one. In both cases, the pair (S, A) is called a

measurable space.

1. Let S be ! (the set of real numbers, or otherwise known as the real

line) and define C0 as follows:

( )( ]( )[

, x , , x , x, , x, , x, y ,

)(

)

{ }

C0 = all intervals in ! = .

( ][ )[ ]

x, y , x, y , x, y ; x, y ! , x < y

12 1 Basic Concepts of Set Theory

it the Borel -field (over the real line). The pair (! , B) is called the Borel real

line.

THEOREM 5 Each one of the following classes generates the Borel -field.

C1 = {(x, y]; x, y !, x < y},

C2 = {[ x, y); x, y ! , x < y},

C5 = {( x, ); x !},

C6 = {[ x, ); x !},

C7 = {( , x ); x !},

C8 = {( , x ]; x !}.

Also the classes C j, j = 1, . . . , 8 generate the Borel -field, where for j = 1, . . . ,

8, Cj is defined the same way as Cj is except that x, y are restricted to the

rational numbers.

(C) (C). Thus, in order to prove the theorem, it suffices to prove that B

(Cj), B (Cj), j = 1, 2, . . . , 8, and in order to prove this, it suffices to show

that C0 (Cj), C0 (Cj), j = 1, 2, . . . , 8. As an example, we show that C0

(C7). Consider xn x. Then (, xn) (C7) and hence In=1 (, xn) (C7).

But

I (, x ) = (, x ].

n

n=1

(x, ) = (, x] , [x, ) = (, x) ,

c c

(x, y) = (, y) (, x] = (, y) (x, ) (C ), 7

7 7

[x, y] = (, y] [x, ) (C ). 7

sequences of rational numbers convergent to given irrationals x, y.

1.1 Some Definitions andExercises

Notation 13

{ } {(, x) (, x ), (, x) (, x ],

C0 = all rectangles in ! 2 =

(, x] (, x ), (, x] (, x ],

(x, ) (x , ), , [x, ) [x , ), ,

(x, y) (x , y), , [x, y] [x , y],

x, y, x , y ! , x < y, x < y}.

Borel -field. A theorem similar to Theorem 5 holds here too.

3. Let S = ! ! ! = ! k (k copies of ! ) and define C0 in a way

similar to that in (2) above. The -field generated by C0 is denoted by Bk and

is the k-dimensional Borel -field. A theorem similar to Theorem 5 holds here

too.

Exercises

1.2.1 Verify that the classes defined in Examples 1, 2 and 3 on page 9 are

fields.

1.2.2 Show that in the definition of a field (Definition 2), property (F3) can

be replaced by (F3) which states that if A1, A2 F, then A1 A2 F.

1.2.3 Show that in Definition 3, property (A3) can be replaced by (A3),

which states that if

Aj A, j = 1, 2, then I Aj A.

j =1

1.2.4 Refer to Example 4 on -fields on page 10 and explain why S was taken

to be uncountable.

1.2.5 Give a formal proof of the fact that the class AA defined in Remark 1

is a -field.

1.2.6 Refer to Definition 1 and show that all three sets A, A and lim An,

n

whenever it exists, belong to A provided An, n 1, belong to A.

1.2.7 Let S = {1, 2, 3, 4} and define the class C of subsets of S as follows:

{ {} {} {} { } { } { } { } { } { }

C = , 1 , 2 , 3 , 4 , 1, 2 , 1, 3 , 1, 4 , 2, 3 , 2, 4 ,

Determine whether or not C is a field.

1.2.8 Complete the proof of the remaining parts in Theorem 5.

14 2 Some Probabilistic Concepts and Results

Chapter 2

and Results

Intuitively by an experiment one pictures a procedure being carried out under

a certain set of conditions whereby the procedure can be repeated any number

of times under the same set of conditions, and upon completion of the proce-

dure certain results are observed. An experiment is a deterministic experiment

if, given the conditions under which the experiment is carried out, the outcome

is completely determined. If, for example, a container of pure water is brought

to a temperature of 100C and 760 mmHg of atmospheric pressure the out-

come is that the water will boil. Also, a certificate of deposit of $1,000 at the

annual rate of 5% will yield $1,050 after one year, and $(1.05)n 1,000 after n

years when the (constant) interest rate is compounded. An experiment for

which the outcome cannot be determined, except that it is known to be one of

a set of possible outcomes, is called a random experiment. Only random

experiments will be considered in this book. Examples of random experiments

are tossing a coin, rolling a die, drawing a card from a standard deck of playing

cards, recording the number of telephone calls which arrive at a telephone

exchange within a specified period of time, counting the number of defective

items produced by a certain manufacturing process within a certain period of

time, recording the heights of individuals in a certain class, etc. The set of all

possible outcomes of a random experiment is called a sample space and is

denoted by S. The elements s of S are called sample points. Certain subsets of

S are called events. Events of the form {s} are called simple events, while an

event containing at least two sample points is called a composite event. S and

are always events, and are called the sure or certain event and the impossible

event, respectively. The class of all events has got to be sufficiently rich in order

to be meaningful. Accordingly, we require that, if A is an event, then so is its

complement Ac. Also, if Aj, j = 1, 2, . . . are events, then so is their union U j Aj.

14

2.1 Probability Functions and Some Basic

2.4 Properties

Combinatorial

and Results 15

(In the terminology of Section 1.2, we require that the events associated with

a sample space form a -field of subsets in that space.) It follows then that I j Aj

is also an event, and so is A1 A2, etc. If the random experiment results in s and

s A, we say that the event A occurs or happens. The U j Aj occurs if at least

one of the Aj occurs, the I j Aj occurs if all Aj occur, A1 A2 occurs if A1 occurs

but A2 does not, etc.

The next basic quantity to be introduced here is that of a probability

function (or of a probability measure).

DEFINITION 1 A probability function denoted by P is a (set) function which assigns to each

event A a number denoted by P(A), called the probability of A, and satisfies

the following requirements:

(P1) P is non-negative; that is, P(A) 0, for every event A.

(P2) P is normed; that is, P(S) = 1.

(P3) P is -additive; that is, for every collection of pairwise (or mutually)

disjoint events Aj, j = 1, 2, . . . , we have P(j Aj) = j P(Aj).

This is the axiomatic (Kolmogorov) definition of probability. The triple

(S, class of events, P) (or (S, A, P)) is known as a probability space.

REMARK 1 If S is finite, then every subset of S is an event (that is, A is taken

to be the discrete -field). In such a case, there are only finitely many events

and hence, in particular, finitely many pairwise disjoint events. Then (P3) is

reduced to:

(P3) P is finitely additive; that is, for every collection of pairwise disjoint

events, Aj, j = 1, 2, . . . , n, we have

n n

( )

P A j = P A j .

j =1 j =1

Actually, in such a case it is sufficient to assume that (P3) holds for any two

disjoint events; (P3) follows then from this assumption by induction.

(C1) P() = 0. In fact, S = S + + ,

so that

( ) ( ) ( ) ( )

P S = P S ++ = P S +P + ,

or

( )

1 = 1+ P + and ( )

P = 0,

since P() 0. (So P() = 0. Any event, possibly " , with probability 0 is

called a null event.)

(C2) P is finitely additive; that is for any event Aj, j = 1, 2, . . . , n such that

Ai Aj = , i j,

16 2 Some Probabilistic Concepts and Results

n n

P A j = P A j .

j =1 j =1

( )

Indeed, for ( ) ( ) ( )

Aj = 0, j n + 1, P jn= 1Aj = P j= 1PAj = j= 1P Aj =

( )

jn= 1P Aj .

(C3) For every event A, P(Ac) = 1 P(A). In fact, since A + Ac = S,

( ) ()

P A + Ac = P S , or ( ) ( )

P A + P Ac = 1,

(C4) P is a non-decreasing function; that is A1 A2 implies P(A1) P(A2).

In fact,

(

A2 = A1 + A2 A1 , )

hence

( ) ( ) (

P A2 = P A1 + P A2 A1 , )

and therefore P(A2) P(A1).

REMARK 2 If A1 A2, then P(A2 A1) = P(A2) P(A1), but this is not true,

in general.

(C5) 0 P(A) 1 for every event A. This follows from (C1), (P2), and (C4).

(C6) For any events A1, A2, P(A1 A2) = P(A1) + P(A2) P(A1 A2).

In fact,

(

A1 A2 = A1 + A2 A1 A2 . )

Hence

( ) ( ) (

P A1 A2 = P A1 + P A2 A1 A2 )

= P( A ) + P( A ) P( A A ),

1 2 1 2

since A1 A2 A2 implies

( ) ( ) (

P A2 A1 A2 = P A2 P A1 A2 . )

(C7) P is subadditive; that is,

P U A j P A j

j =1 j =1

( )

and also

n n

P U A j P A j .

j =1 j =1

( )

2.1 Probability Functions and Some Basic

2.4 Properties

Combinatorial

and Results 17

j =1

n

j =1

A special case of a probability space is the following: Let S = {s1, s2, . . . , sn},

let the class of events be the class of all subsets of S, and define P as P({sj}) =

1/n, j = 1, 2, . . . , n. With this definition, P clearly satisfies (P1)(P3) and this

is the classical definition of probability. Such a probability function is called a

uniform probability function. This definition is adequate as long as S is finite

and the simple events {sj}, j = 1, 2, . . . , n, may be assumed to be equally

likely, but it breaks down if either of these two conditions is not satisfied.

However, this classical definition together with the following relative frequency

(or statistical) definition of probability served as a motivation for arriving at

the axioms (P1)(P3) in the Kolmogorov definition of probability. The relative

frequency definition of probability is this: Let S be any sample space, finite

or not, supplied with a class of events A. A random experiment associated with

the sample space S is carried out n times. Let n(A) be the number of times that

the event A occurs. If, as n , lim[n(A)/n] exists, it is called the probability

of A, and is denoted by P(A). Clearly, this definition satisfies (P1), (P2) and

(P3).

Neither the classical definition nor the relative frequency definition of

probability is adequate for a deep study of probability theory. The relative

frequency definition of probability provides, however, an intuitively satisfac-

tory interpretation of the concept of probability.

We now state and prove some general theorems about probability

functions.

n

( )

n

( )

P U A j = P A j P A j A j

j =1 j =1 1 j < j n

1 2

1 2

+

1 j1 < j2 < j3 n

(

P Aj Aj Aj

1 2 3

)

( ) ( )

n+1

+ 1 P A1 A2 An .

have proven the case n = 2 as consequence (C6) of the definition of probability

functions. Now assume the result to be true for n = k, and prove it for

n = k + 1.

18 2 Some Probabilistic Concepts and Results

We have

k+1 k

P U A j = P U A j Ak+1

j =1 j =1

k k

j =1 j =1

(

= P U A j + P Ak+1 P U A j Ak+1

)

k

( )

= P A j P A j1 A j2

j = 1 1 j1 < j2 k

( )

+ P A j1 A j2 A j3

1 j1 < j2 < j3 k

( )

k

( ) (

P A1 A2 Ak + P Ak+1 P U A j Ak+1 ) ( ) ( )

k+1

+ 1

j =1

k+1

= P Aj

j =1

( ) P( A A ) 1 ji < j2 k

j1 j2

+ P( A A A ) j1 j2 j3

1 ji < j2 < j3 k

k

( ) ( ) ( ) .

k+1

+ 1 P A1 A2 Ak P U A j Ak+1 (1)

But j =1

k

( )

k

( )

P U A j Ak+1 = P A j Ak+1 P A j1 A j2 Ak+1

j =1 j =1 1 j1 < j2 k

( )

+

1 j1 < j2 < j3 k

(

P A j1 A j2 A j3 Ak+1 )

( ) ( )

k

+ 1 P A j1 A jk 11 Ak+1

1 j1 < j2 jk 1 k

( ) ( )

k+1

+ 1 P A1 Ak Ak+1 .

k+1 k+1

( )

k

( )

P U A j = P A j P A j A j + P A j Ak+1

j =1 j =1 1 j < j k j =1

1 2

( )

1 2

(

+ P A j A j A j + P A j A j Ak+1

1 j < j < j k 1 j < j k

1 2 3

) ( 1 2

)

1 2 3 1 2

( ) [ P( A A )

k+1

+ 1 1 k

+

1 j1 < j2 < < jk 1 k

(

P A j A j Ak+1

1

k 1

)

2.1 Probability Functions and Some Basic

2.4 Properties

Combinatorial

and Results 19

( ) P( A A A )

k+ 2

+ 1 1 k k+1

k+1

= P( A ) P( A A )

j j1 j2

j=1 1 j1 < j2 k+1

+ P( A A A ) j1 j2 j3

1 j1 < j2 < j3 k+1

( ) ( )

k+ 2

+ 1 P A1 Ak+1 .

(

P lim An = lim P An .

n

) n

( )

PROOF Let us first assume that An . Then

lim An = U A j.

n

j=1

We recall that

( ) ( )

j =1

(

= A1 + A2 A1 + A3 A2 + , ) ( )

by the assumption that An . Hence

(

)

P lim An = P U A j = P A1 + P A2 A1

n j =1

( ) ( )

( ( ) )

+ P A3 A2 + + P An An1 +

= lim[ P( A ) + P( A A ) + + P( A A )]

1 2 1 n n1

n

= lim[ P( A ) + P( A ) P( A )

1 2 1

n

+ P( A ) P( A ) + + P( A ) P( A )]

3 2 n n1

= lim P( A ). n

n

Thus

(

P lim An = lim P An .

n

) n

( )

Now let An . Then Acn , so that

lim Anc = U Acj .

n

j=1

20 2 Some Probabilistic Concepts and Results

Hence

(

n

)

j=1 n

( )

P lim Anc = P U Acj = lim P Anc ,

or equivalently,

c

j = 1 n [ ( )]

P I A j = lim 1 P An , or 1 P I A j = 1 lim P An .

j =1 n

( )

Thus

n

( )

(

lim P An = P I A j = P lim An ,

j =1 n

)

and the theorem is established.

This theorem will prove very useful in many parts of this book.

Exercises

2.1.1 If the events Aj, j = 1, 2, 3 are such that A1 A2 A3 and P(A1) = 1

4

,

P(A2) = 125 , P(A3) = 7 , compute the probability of the following events:

12

1

c

2.1.2 If two fair dice are rolled once, what is the probability that the total

number of spots shown is

i) Equal to 5?

ii) Divisible by 3?

2.1.3 Twenty balls numbered from 1 to 20 are mixed in an urn and two balls

are drawn successively and without replacement. If x1 and x2 are the numbers

written on the first and second ball drawn, respectively, what is the probability

that:

i) x1 + x2 = 8?

ii) x1 + x2 5?

2.1.4 Let S = {x integer; 1 x 200} and define the events A, B, and C by:

{

A = x S ; x is divisible by 7 , }

B = {x S ; x = 3n + 10 }

for some positive integer n ,

{

C = x S ; x 2 + 1 375 . }

2.2

2.4 Conditional

Combinatorial

Probability

Results 21

Compute P(A), P(B), P(C), where P is the equally likely probability function

on the events of S.

2.1.5 Let S be the set of all outcomes when flipping a fair coin four times and

let P be the uniform probability function on the events of S. Define the events

A, B as follows:

{ }

A = s S ; s contains more T s than H s ,

B = {s S ; any T in s precedes every H in s}.

2.1.6 Suppose that the events Aj, j = 1, 2, . . . are such that

P( A j ) < .

j=1

) = 0.

that P(A

2.1.7 Consider the events Aj, j = 1, 2, . . . and use Definition 1 in Chapter 1

and Theorem 2 herein in order to show that

( ) ( ) ( )

P A lim inf P An lim sup P An P A .

n n

( )

In this section, we shall introduce the concepts of conditional probability and

stochastic independence. Before the formal definition of conditional probabil-

ity is given, we shall attempt to provide some intuitive motivation for it. To this

end, consider a balanced die and suppose that the sides bearing the numbers

1, 4 and 6 are painted red, whereas the remaining three sides are painted black.

The die is rolled once and we are asked for the probability that the upward

side is the one bearing the number 6. Assuming the uniform probability

function, the answer is, clearly, 16 . Next, suppose that the die is rolled once as

before and all that we can observe is the color of the upward side but not the

number on it (for example, we may be observing the die from a considerable

distance, so that the color is visible but not the numbers on the sides). The

same question as above is asked, namely, what is the probability that the

number on the uppermost side is 6. Again, by assuming the uniform probabil-

ity function, the answer now is 13 . This latter probability is called the condi-

tional probability of the number 6 turning up, given the information that the

uppermost side was painted red. Letting B stand for the event that number 6

appears and A for the event that the uppermost side is red, the above-

22 2 Some Probabilistic Concepts and Results

this is equal to the quotient P(A B)/P(A). Or suppose that, for the purposes

of a certain study, we observe two-children families in a certain locality, and

record the gender of the children. A sample space for this experiment is the

following: S = {bb, bg, gb, gg}, where b stands for boy and g for girl, and bg, for

example, indicates that the boy is older than the girl. Suppose further

(although this is not exactly correct) that: P({bb}) = P({bg}) = P({gb}) = P({gg})

= 14 , and define the events A and B as follows: A = children of one gender =

{bb, gg}, B = at least one boy = {bb, bg, gb}. Then P(A|B) = P(A B)/

P(B) = 13 .

From these and other examples, one is led to the following definition of

conditional probability.

DEFINITION 2 Let A be an event such that P(A) > 0. Then the conditional probability, given

A, is the (set) function denoted by P(|A) and defined for every event B as

follows:

( ) (P( A) ) .

P A B

P BA =

The set function P(|A) is actually a probability function. To see this, it

suffices to prove the P(|A) satisfies (P1)(P3). We have: P(B|A) # 0 for every

event B, clearly. Next,

P SA P A

PSA =

(

P j = 1 A j A P j = 1 A j A

)

P A j A = =

j =1 P A ( ) P A ( )

(

P Aj A )=

P( A )

P( A )

1

= A = A.

( )

P A j =1

j

j =1 ( )

P A j =1

j

the intersection of a finite number of events.

n1

P I A j > 0.

j =1

Then

2.4

2.2 Conditional

Combinatorial

Probability

Results 23

n

(

P I A j = P An A1 A2 An1

j =1

)

(

P An1 A1 An 2 P A2 A1 P A1 . ) ( )( )

(The proof of this theorem is left as an exercise; see Exercise 2.2.4.)

REMARK 3 The value of the above formula lies in the fact that, in general, it

is easier to calculate the conditional probabilities on the right-hand side. This

point is illustrated by the following simple example.

EXAMPLE 1 An urn contains 10 identical balls (except for color) of which five are black,

three are red and two are white. Four balls are drawn without replacement.

Find the probability that the first ball is black, the second red, the third white

and the fourth black.

Let A1 be the event that the first ball is black, A2 be the event that the

second ball is red, A3 be the event that the third ball is white and A4 be the

event that the fourth ball is black. Then

(

P A1 A2 A3 A4 )

( )(

= P A4 A1 A2 A3 P A3 A1 A2 P A2 A1 P A1 , )( )( )

and by using the uniform probability function, we have

( )

P A1 =

5

10

, (

P A2 A1 = ) 3

9

, (

P A3 A1 A2 = ) 2

8

,

(

P A4 A1 A2 A3 = ) 4

7

.

42

! 0.0238.

Now let Aj, j = 1, 2, . . . , be events such that Ai Aj = , i j, and j Aj =

S. Such a collection of events is called a partition of S. The partition is finite

or (denumerably) infinite, accordingly, as the events Aj are finitely or

(denumerably) infinitely many. For any event, we clearly have:

(

B = B Aj .

j

)

Hence

( ) (

P B = P B Aj = P B Aj P Aj ,

j

) j

( )( )

provided P(Aj) > 0, for all j. Thus we have the following theorem.

THEOREM 4 (Total Probability Theorem) Let {Aj, j = 1, 2, . . . } be a partition of S with

P(Aj) > 0, all j. Then for B A, we have P(B) = jP(B|Aj)P(Aj).

This formula gives a way of evaluating P(B) in terms of P(B|Aj) and

P(Aj), j = 1, 2, . . . . Under the condition that P(B) > 0, the above formula

24 2 Some Probabilistic Concepts and Results

fact,

P A B P B A P( A ) P B A P( A )

(

P Aj B =

j j j j j

i i i

Thus

THEOREM 5 (Bayes Formula) If {Aj, j = 1, 2, . . .} is a partition of S and P(Aj) > 0, j = 1,

2, . . . , and if P(B) > 0, then

(

P Aj B = ) (

)( ).

P B Aj P Aj

P( B A ) P( A )

i i i

{Aj, j 1} forms a partition of S, as only then are the above theorems true.

The following simple example serves as an illustration of Theorems 4

and 5.

EXAMPLE 2 A multiple choice test question lists five alternative answers, of which only one

is correct. If a student has done the homework, then he/she is certain to

identify the correct answer; otherwise he/she chooses an answer at random.

Let p denote the probability of the event A that the student does the home-

work and let B be the event that he/she answers the question correctly. Find

the expression of the conditional probability P(A|B) in terms of p.

By noting that A and Ac form a partition of the appropriate sample space,

an application of Theorems 4 and 5 gives

( )

P AB =

( )( )

P BA P A

=

1 p

=

5p

.

( )( ) (

P BA P A +P BA P A c

)( ) c 1

(

1 p + 1 p

5

4p+1

)

Furthermore, it is easily seen that P(A|B) = P(A) if and only if p = 0 or 1.

For example, for p = 0.7, 0.5, 0.3, we find, respectively, that P(A|B) is

approximately equal to: 0.92, 0.83 and 0.68.

Of course, there is no reason to restrict ourselves to one partition of S

only. We may consider, for example, two partitions {Ai, i = 1, 2, . . .} {Bj, j = 1,

2, . . . }. Then, clearly,

(

Ai = Ai B j ,

j

) i = 1, 2, ,

Bj = ( B A ), j i j = 1, 2, ,

i

and

{A B , i = 1, 2, , j = 1, 2, }

i j

2.4 Combinatorial

Exercises

Results 25

is a partition of S. In fact,

(A B ) (A

i j i )

B j = if (i, j ) (i, j )

and

( A B ) = ( A B ) = A = S.

i j i j i

i, j i j i

The expression P(Ai Bj) is called the joint probability of Ai and Bj. On the

other hand, from

Ai = Ai Bj

j

( ) and Bj = Ai Bj ,

i

( )

we get

( ) (

P Ai = P Ai B j = P Ai B j P B j ,

j

) j

( )( )

provided P(Bj) > 0, j = 1, 2, . . . , and

( ) (

P B j = P Ai B j = P B j Ai P Ai ,

i

) i

( )( )

provided P(Ai) > 0, i = 1, 2, . . . . The probabilities P(Ai), P(Bj) are called

marginal probabilities. We have analogous expressions for the case of more

than two partitions of S.

Exercises

2.2.1 If P(A|B) > P(A), then show that P(B|A) > P(B) (P(A)P(B) > 0).

2.2.2 Show that:

i) P(Ac|B) = 1 P(A|B);

ii) P(A B|C) = P(A|C) + P(B|C) P(A B|C).

Also show, by means of counterexamples, that the following equations need

not be true:

iii) P(A|Bc) = 1 P(A|B);

iv) P(C|A + B) = P(C|A) + P(C|B).

2.2.3 If A B = and P(A + B) > 0, express the probabilities P(A|A + B)

and P(B|A + B) in terms of P(A) and P(B).

2.2.4 Use induction to prove Theorem 3.

2.2.5 Suppose that a multiple choice test lists n alternative answers of which

only one is correct. Let p, A and B be defined as in Example 2 and find Pn(A|B)

26 2 Some Probabilistic Concepts and Results

in terms of n and p. Next show that if p is fixed but different from 0 and 1, then

Pn(A|B) increases as n increases. Does this result seem reasonable?

2.2.6 If Aj, j = 1, 2, 3 are any events in S, show that {A1, Ac1 A2, Ac1 Ac2

A3, (A1 A2 A3)c} is a partition of S.

2.2.7 Let {Aj, j = 1, . . . , 5} be a partition of S and suppose that P(Aj) = j/15

and P(A|Aj) = (5 j)/15, j = 1, . . . , 5. Compute the probabilities P(Aj|A),

j = 1, . . . , 5.

2.2.8 A girls club has on its membership rolls the names of 50 girls with the

following descriptions:

20 blondes, 15 with blue eyes and 5 with brown eyes;

25 brunettes, 5 with blue eyes and 20 with brown eyes;

5 redheads, 1 with blue eyes and 4 with green eyes.

If one arranges a blind date with a club member, what is the probability that:

i) The girl is blonde?

ii) The girl is blonde, if it was only revealed that she has blue eyes?

2.2.9 Suppose that the probability that both of a pair of twins are boys is 0.30

and that the probability that they are both girls is 0.26. Given that the probabil-

ity of a child being a boy is 0.52, what is the probability that:

i) The second twin is a boy, given that the first is a boy?

ii) The second twin is a girl, given that the first is a girl?

2.2.10 Three machines I, II and III manufacture 30%, 30% and 40%, respec-

tively, of the total output of certain items. Of them, 4%, 3% and 2%, respec-

tively, are defective. One item is drawn at random, tested and found to be

defective. What is the probability that the item was manufactured by each one

of the machines I, II and III?

2.2.11 A shipment of 20 TV tubes contains 16 good tubes and 4 defective

tubes. Three tubes are chosen at random and tested successively. What is the

probability that:

i) The third tube is good, if the first two were found to be good?

ii) The third tube is defective, if one of the other two was found to be good

and the other one was found to be defective?

2.2.12 Suppose that a test for diagnosing a certain heart disease is 95%

accurate when applied to both those who have the disease and those who do

not. If it is known that 5 of 1,000 in a certain population have the disease in

question, compute the probability that a patient actually has the disease if the

test indicates that he does. (Interpret the answer by intuitive reasoning.)

2.2.13 Consider two urns Uj, j = 1, 2, such that urn Uj contains mj white balls

and nj black balls. A ball is drawn at random from each one of the two urns and

2.4 Combinatorial

2.3 Independence

Results 27

is placed into a third urn. Then a ball is drawn at random from the third urn.

Compute the probability that the ball is black.

2.2.14 Consider the urns of Exercise 2.2.13. A balanced die is rolled and if

an even number appears, a ball, chosen at random from U1, is transferred to

urn U2. If an odd number appears, a ball, chosen at random from urn U2, is

transferred to urn U1. What is the probability that, after the above experiment

is performed twice, the number of white balls in the urn U2 remains the same?

2.2.15 Consider three urns Uj, j = 1, 2, 3 such that urn Uj contains mj white

balls and nj black balls. A ball, chosen at random, is transferred from urn U1 to

urn U2 (color unnoticed), and then a ball, chosen at random, is transferred

from urn U2 to urn U3 (color unnoticed). Finally, a ball is drawn at random

from urn U3. What is the probability that the ball is white?

2.2.16 Consider the urns of Exercise 2.2.15. One urn is chosen at random

and one ball is drawn from it also at random. If the ball drawn was white, what

is the probability that the urn chosen was urn U1 or U2?

2.2.17 Consider six urns Uj, j = 1, . . . , 6 such that urn Uj contains mj ( 2)

white balls and nj ( 2) black balls. A balanced die is tossed once and if the

number j appears on the die, two balls are selected at random from urn Uj.

Compute the probability that one ball is white and one ball is black.

2.2.18 Consider k urns Uj, j = 1, . . . , k each of which contain m white balls

and n black balls. A ball is drawn at random from urn U1 and is placed in urn

U2. Then a ball is drawn at random from urn U2 and is placed in urn U3 etc.

Finally, a ball is chosen at random from urn Uk1 and is placed in urn Uk. A ball

is then drawn at random from urn Uk. Compute the probability that this last

ball is black.

2.3 Independence

For any events A, B with P(A) > 0, we defined P(B|A) = P(A B)/P(A). Now

P(B|A) may be >P(B), < P(B), or = P(B). As an illustration, consider an urn

containing 10 balls, seven of which are red, the remaining three being black.

Except for color, the balls are identical. Suppose that two balls are drawn

successively and without replacement. Then (assuming throughout the uni-

form probability function) the conditional probability that the second ball is

red, given that the first ball was red, is 69 , whereas the conditional probability

that the second ball is red, given that the first was black, is 79 . Without any

knowledge regarding the first ball, the probability that the second ball is red is

7

10

. On the other hand, if the balls are drawn with replacement, the probability

that the second ball is red, given that the first ball was red, is 107 . This probabil-

ity is the same even if the first ball was black. In other words, knowledge of the

event which occurred in the first drawing provides no additional information in

28 2 Some Probabilistic Concepts and Results

calculating the probability of the event that the second ball is red. Events like

these are said to be independent.

As another example, revisit the two-children families example considered

earlier, and define the events A and B as follows: A = children of both

genders, B = older child is a boy. Then P(A) = P(B) = P(B|A) = 12 . Again

knowledge of the event A provides no additional information in calculating

the probability of the event B. Thus A and B are independent.

More generally, let A, B be events with P(A) > 0. Then if P(B|A) = P(B),

we say that the even B is (statistically or stochastically or in the probability

sense) independent of the event A. If P(B) is also > 0, then it is easily seen that

A is also independent of B. In fact,

P A B P B A P( A) P B P A

P AB =

That is, if P(A), P(B) > 0, and one of the events is independent of the other,

then this second event is also independent of the first. Thus, independence is

a symmetric relation, and we may simply say that A and B are independent. In

this case P(A B) = P(A)P(B) and we may take this relationship as the

definition of independence of A and B. That is,

DEFINITION 3 The events A, B are said to be (statistically or stochastically or in the probabil-

ity sense) independent if P(A B) = P(A)P(B).

Notice that this relationship is true even if one or both of P(A), P(B) = 0.

As was pointed out in connection with the examples discussed above,

independence of two events simply means that knowledge of the occurrence of

one of them helps in no way in re-evaluating the probability that the other

event happens. This is true for any two independent events A and B, as follows

from the equation P(A|B) = P(A), provided P(B) > 0, or P(B|A) = P(B),

provided P(A) > 0. Events which are intuitively independent arise, for exam-

ple, in connection with the descriptive experiments of successively drawing

balls with replacement from the same urn with always the same content, or

drawing cards with replacement from the same deck of playing cards, or

repeatedly tossing the same or different coins, etc.

What actually happens in practice is to consider events which are inde-

pendent in the intuitive sense, and then define the probability function P

appropriately to reflect this independence.

The definition of independence generalizes to any finite number of events.

Thus:

DEFINITION 4 The events Aj, j = 1, 2, . . . , n are said to be (mutually or completely) indepen-

dent if the following relationships hold:

( 1 k

) ( )

P Aj Aj = P Aj P Aj

1

( ) k

n. These events are said to be pairwise independent if P(Ai Aj) = P(Ai)P(Aj)

for all i j.

2.4 Combinatorial

2.3 Independence

Results 29

then they are pairwise independent. The converse need not be true, as the

example below illustrates. Also there are

n n n n n

+ + + =2 =2 n1

n n

2 3 n 1 0

all necessary. For example, for n = 3 we will have:

( ) ( ) ( ) ( )

P A1 A2 A3 = P A1 P A2 P A3 ,

P( A A ) = P( A )P( A ),

1 2 1 2

P( A A ) = P( A )P( A ),

1 3 1 3

P( A A ) = P( A )P( A ).

2 3 2 3

That these four relations are necessary for the characterization of indepen-

dence of A1, A2, A3 is illustrated by the following examples:

Let S = {1, 2, 3, 4}, P({1}) = = P({4}) = 14 , and set A1 = {1, 2}, A2 = {1, 3},

A3 = {1, 4}. Then

A1 A2 = A1 A3 = A2 A3 = 1 , and {} A1 A2 A3 = 1 .{}

Thus

( ) ( ) (

P A1 A2 = P A1 A3 = P A2 A3 = P A1 A2 A3 = ) ( ) 1

4

.

Next,

(

P A1 A2 =

1 1 1

)

= = P A1 P A2 ,

4 2 2

( ) ( )

( 1 1 1

)

P A1 A3 = = = P A1 P A3 ,

4 2 2

( ) ( )

( 1 1 1

)

P A1 A3 = = = P A2 P A3 ,

4 2 2

( ) ( )

but

(

P A1 A2 A3 = ) 1 1 1 1

( ) ( ) ( )

= P A1 P A2 P A3 .

4 2 2 2

({ })

P 1 =

1

8

, P 2 ({ }) = P({3}) = P({4}) = 163 , P({5}) = 165 .

30 2 Some Probabilistic Concepts and Results

Let

{ } {

A1 = 1, 2, 3 , A2 = 1, 2, 4 , A3 = 1, 3, 4 . } { }

Then

A1 A2 = 1, 2 , { } A1 A2 A3 = 1 . {}

Thus

(

P A1 A2 A3 = ) 1 1 1 1

= = P A1 P A2 P A3 ,

8 2 2 2

( ) ( ) ( )

but

(

P A1 A2 = ) 5 1 1

= P A1 P A2 .

16 2 2

( ) ( )

The following result, regarding independence of events, is often used by

many authors without any reference to it. It is the theorem below.

THEOREM 6 If the events A1, . . . , An are independent, so are the events A1, . . . , An, where

Aj is either Aj or Acj, j = 1, . . . , n.

PROOF The proof is done by (a double) induction. For n = 2, we have to

show that P(A1 A2) = P(A1)P(A2). Indeed, let A1 = A1 and A2 = Ac2. Then

P(A1 A2) = P(A1 Ac2) = P[A1 (S A2)] = P(A1 A1 A2) = P(A1)

P(A1 A2) = P(A1) P(A1)P(A2) = P(A1)[1 P(A2)] = P(A1)P(Ac2) = P(A1)P(A2).

Similarly if A1 = Ac1 and A2 = A2. For A1 = Ac1 and A2 = Ac2, P(A1 A2) =

P(Ac1 Ac2) = P[(S A1) Ac2] = P(Ac2 A1 Ac2) = P(Ac2) P(A1 Ac2) = P(Ac2)

P(A1)P(Ac2) = P(Ac2)[1 P(A1)] = P(Ac2)P(Ac1) = P(A1)P(A2).

Next, assume the assertion to be true for k events and show it to be true

for k + 1 events. That is, we suppose that P(A1 Ak) = P(A1 ) P(Ak),

and we shall show that P(A1 Ak+1) = P(A1 ) P(Ak+1). First, assume

that Ak+1 = Ak+1, and we have to show that

( ) (

P A1 Ak +1 = P A1 Ak Ak+1 )

= P( A) P( A )P( A ).

1 k k+1

Ai = Ai, i = 2, . . . , k, then

( ) [(

P A1c A2 Ak Ak+1 = P S A1 A2 Ak Ak+1 ) ]

(

= P A2 Ak Ak+1 A1 A2 Ak Ak+1 )

= P( A A A ) P( A A A A )

2 k k+1 1 2 k k+1

= P( A ) P( A ) P( A ) P( A ) P( A ) P( A ) P( A )

2 k k+1 1 2 k k+1

2 k k+1 1 1

c

2 k k+1

2.4 Combinatorial

2.3 Independence

Results 31

This is, clearly, true if Ac1 is replaced by any other Aci, i = 2, . . . , k. Now, for

! < k, assume that

(

P A1c Alc Al+1 Ak +1 )

( ) ( )( )

= P A1c P Alc P Al+1 P Ak +1 ( )

and show that

(

P A1c Alc Alc+1 Al+ 2 Ak +1 )

=P A ( ) P( A )P( A )P( A ) P( A ).

c

1

c

l

c

l +1 l+ 2 k +1

Indeed,

(

P A1c Alc Alc+1 Al+ 2 Ak+1 )

[ (

= P A1c Alc S Al+1 Al+ 2 Ak+1 ) ]

(

= P A A Al+ 2 Ak+1

1

c c

l

(

= P A A Al+ 2 Ak+1

1

c c

l )

(

P A A Al+1 Al+ 2 Ak+1

1

c c

l )

( ) ( )( ) ( )

= P A1c P Alc P Al+ 2 P Ak+1

P( A ) P( A ) P( A ) P( A ) P( A )

1

c c

l l +1 l+ 2 k+1

= P( A ) P( A )P( A ) P( A )[1 P( A )]

1

c c

l l+ 2 k+1 l +1

= P( A ) P( A ) P( A ) P( A ) P( A )

1

c c

l l+ 2 k+1

c

l +1

= P( A ) P( A )P( A )P( A ) P( A ),

1

c c

l

c

l +1 l+ 2 k+1

as was to be seen. It is also, clearly, true that the same result holds if the ! Ai s

which are Aci are chosen in any one of the (k!) possible ways of choosing ! out

of k. Thus, we have shown that

(

P A1 Ak Ak+1 = P A1 P Ak P Ak+1 . ) ( ) ( ) ( )

Finally, under the assumption that

( )

P A1 Ak = P A1 P Ak , ( ) ( )

take Ak+1 = A , and show that

c

k+1

(

P A1 Ak Akc+1 = P A1 P Ak P Akc+1 .) ( ) ( ) ( )

32 2 Some Probabilistic Concepts and Results

In fact,

( ) [(

P A1 Ak Akc+1 = P A1 Ak S Ak+1 ( ))]

(

= P A1 Ak A1 Ak Ak+1 )

= P( A A ) P( A A A )

1 k 1 k k+1

= P( A) P( A ) P( A) P( A )P( A )

1 k 1 k k+1

= P( A) P( A )[1 P( A )]

1 k k+1

= P( A) P( A )P( A ).

1 k

c

k+1

Now, for j = 1, 2, let Ej be an experiment having the sample space Sj. One

may look at the pair (E1, E2) of experiments, and then the question arises as to

what is the appropriate sample space for this composite or compound experi-

ment, also denoted by E1 E2. If S stands for this sample space, then, clearly,

S = S1 S2 = {(s1, s2); s1 S1, s2 S2}. The corresponding events are, of course,

subsets of S. The notion of independence also carries over to experiments.

Thus, we say that the experiments E1 and E2 are independent if P(B1 B2) =

P(B1)P(B2) for all events B1 associated with E1 alone, and all events B2 associ-

ated E2 alone.

What actually happens in practice is to start out with two experiments E1,

E2 which are intuitively independent, such as the descriptive experiments (also

mentioned above) of successively drawing balls with replacement from the

same urn with always the same content, or drawing cards with replacement

from the same deck of playing cards, or repeatedly tossing the same or differ-

ent coins etc., and have the corresponding probability spaces (S1, class of

events, P1) and (S2, class of events, P2), and then define the probability func-

tion P, in terms of P1 and P2, on the class of events in the space S1 S2 so that

it reflects the intuitive independence.

The above definitions generalize in a straightforward manner to any finite

number of experiments. Thus, if Ej, j = 1, 2, . . . , n, are n experiments with

corresponding sample spaces Sj and probability functions Pj on the respective

classes of events, then the compound experiment

(E , E , , E ) = E

1 2 n 1 E2 En

has sample space S, where

S = S1 S n = {(s , , s ); s S , j = 1, 2, , n}.

1 n j j

The class of events are subsets of S, and the experiments are said to be

independent if for all events Bj associated with experiment Ej alone, j = 1,

2, . . . , n, it holds

2.4 Combinatorial

Exercises

Results 33

( ) ( ) ( )

P B1 Bn = P B1 P Bn .

Again, the probability function P is defined, in terms of Pj, j = 1, 2, . . . , n, on

the class of events in S so that to reflect the intuitive independence of the

experiments Ej, j = 1, 2, . . . , n.

In closing this section, we mention that events and experiments which are

not independent are said to be dependent.

Exercises

2.3.1 If A and B are disjoint events, then show that A and B are independent

if and only if at least one of P(A), P(B) is zero.

2.3.2 Show that if the event A is independent of itself, then P(A) = 0 or 1.

2.3.3 If A, B are independent, A, C are independent and B C = , then A,

B + C are independent. Show, by means of a counterexample, that the conclu-

sion need not be true if B C .

2.3.4 For each j = 1, . . . , n, suppose that the events A1, . . . , Am, Bj are

independent and that Bi Bj = , i j. Then show that the events A1, . . . , Am,

nj= 1Bj are independent.

2.3.5 If Aj, j = 1, . . . , n are independent events, show that

n

( )

n

P U A j = 1 P Acj .

j =1 j =1

2.3.6 Jim takes the written and road drivers license tests repeatedly until he

passes them. Given that the probability that he passes the written test is 0.9

and the road test is 0.6 and that tests are independent of each other, what is the

probability that he will pass both tests on his nth attempt? (Assume that

the road test cannot be taken unless he passes the written test, and that once

he passes the written test he does not have to take it again, no matter whether

he passes or fails his next road test. Also, the written and the road tests are

considered distinct attempts.)

2.3.7 The probability that a missile fired against a target is not intercepted by

an antimissile missile is 32 . Given that the missile has not been intercepted, the

probability of a successful hit is 34 . If four missiles are fired independently,

what is the probability that:

i) All will successfully hit the target?

ii) At least one will do so?

How many missiles should be fired, so that:

34 2 Some Probabilistic Concepts and Results

iv) At least one successfully hits its target with probability 0.99?

2.3.8 Two fair dice are rolled repeatedly and independently. The first time a

total of 10 appears, player A wins, while the first time that a total of 6 appears,

player B wins, and the game is terminated. Compute the probabilities that:

i) The game terminates on the nth throw and player A wins;

ii) The same for player B;

iii) Player A wins;

iv) Player B wins;

v) Does the game terminate ever?

2.3.9 Electric current is transmitted from point A to point B provided at

least one of the circuits #1 through #n below is closed. If the circuits close

independently of each other and with respective probabilities pi, i = 1, . . . , n,

determine the probability that:

i) Exactly one circuit is closed;

ii) At least one circuit is closed;

iii) Exactly m circuits are closed for 0 m n;

iv) At least m circuits are closed with m as in part (iii);

v) What do parts (i)(iv) become for p1 = = pn = p, say?

A B

In this section, we will restrict ourselves to finite sample spaces and uniform

probability functions. Some combinatorial results will be needed and we pro-

ceed to derive them here. Also examples illustrating the theorems of previous

sections will be presented.

2.4 Combinatorial Results 35

forms the backbone of the results in this section.

THEOREM 7 Let a task T be completed by carrying out all of the subtasks Tj, j = 1, 2, . . . ,

k, and let it be possible to perform the subtask Tj in nj (different) ways, j = 1,

2, . . . , k. Then the total number of ways the task T may be performed is given

by kj = 1 n j .

PROOF The assertion is true for k = 2, since by combining each one of the n1

ways of performing subtask T1 with each one of the n2 ways of performing

subtask T2, we obtain n1n2 as the total number of ways of performing task T.

Next, assume the result to be true for k = m and establish it for k = m + 1. The

reasoning is the same as in the step just completed, since by combining each

one of the mj = 1 n j ways of performing the first m subtasks with each one of

( )

nm+1 ways of performing substask Tm+1, we obtain mj = 1 n j nm+1 = jm=+11 n j for

the total number of ways of completing task T.

The following examples serve as an illustration to Theorem 7.

EXAMPLE 3 i) A man has five suits, three pairs of shoes and two hats. Then the number

of different ways he can attire himself is 5 3 2 = 30.

ii) Consider the set S = {1, . . . , N} and suppose that we are interested in

finding the number of its subsets. In forming a subset, we consider for each

element whether to include it or not. Then the required number is equal to

the following product of N factors 2 2 = 2N.

iii) Let nj = n(Sj) be the number of points of the sample space Sj, j = 1, 2, . . . ,

k. Then the sample space S = S1 Sk has n(S) = n1 nk sample

points. Or, if nj is the number of outcomes of the experiment Ej, j = 1,

2, . . . , k, then the number of outcomes of the compound experiment E1

. . . Ek is n1 . . . nk.

In the following, we shall consider the problems of selecting balls from an

urn and also placing balls into cells which serve as general models of

many interesting real life problems. The main results will be formulated as

theorems and their proofs will be applications of the Fundamental Principle of

Counting.

Consider an urn which contains n numbered (distinct, but otherwise iden-

tical) balls. If k balls are drawn from the urn, we say that a sample of size k was

drawn. The sample is ordered if the order in which the balls are drawn is taken

into consideration and unordered otherwise. Then we have the following

result.

THEOREM 8 i) The number of ordered samples of size k is n(n 1) (n k + 1) = Pn,k

(permutations of k objects out of n, and in particular, if k = n, Pn,n = 1 2

n = n!), provided the sampling is done without replacement; and is equal to

nk if the sampling is done with replacement.

ii) The number of unordered samples of size k is

36 2 Some Probabilistic Concepts and Results

Pn ,k n n!

= C n ,k = =

k! (

k k! n k !)

if the sampling is done without replacement; and is equal to

n + k 1

( )

N n, k =

k

PROOF

i) The first part follows from Theorem 7 by taking nj = (n j + 1), j = 1, . . . ,

k, and the second part follows from the same theorem by taking nj = n,

j = 1, . . . , k.

ii) For the first part, we have that, if order counts, this number is Pn,k. Since for

every sample of size k one can form k! ordered samples of the same size, if

x is the required number, then Pn,k = xk!. Hence the desired result.

The proof of the second part may be carried out by an appropriate induc-

tion method. However, we choose to present the following short alternative

proof which is due to S. W. Golomb and appeared in the American Mathemati-

cal Monthly, 75, 1968, p. 530. For clarity, consider the n balls to be cards

numbered from 1 to n and adjoin k 1 extra cards numbered from n + 1 to

n + k 1 and bearing the respective instructions: repeat lowest numbered

card, repeat 2nd lowest numbered card, . . . , repeat (k 1)st lowest num-

bered card. Then a sample of size k without replacement from this enlarged

(n + k 1)-card deck corresponds uniquely to a sample of size k from the

original deck with replacement. (That is, take k out of n + k 1, without

replacement so that there will be at least one out of 1, 2, . . . , n, and then apply

the instructions.) Thus, by the first part, the required number is

n + k 1

k

(

= N n, k , )

as was to be seen.

For the sake of illustration of Theorem 8, let us consider the following

examples.

EXAMPLE 4 i(i) Form all possible three digit numbers by using the numbers 1, 2, 3, 4, 5.

(ii) Find the number of all subsets of the set S = {1, . . . , N}.

In part (i), clearly, the order in which the numbers are selected is relevant.

Then the required number is P5,3 = 5 4 3 = 60 without repetitions, and 53 = 125

with repetitions.

In part (ii) the order is, clearly, irrelevant and the required number is (N0)

+ ( 1) + + (NN) = 2N, as already found in Example 3.

N

2.4 Combinatorial Results 37

EXAMPLE 5 An urn contains 8 balls numbered 1 to 8. Four balls are drawn. What is the

probability that the smallest number is 3?

Assuming the uniform probability function, the required probabilities are

as follows for the respective four possible sampling cases:

5

3 1

Order does not count/replacements not allowed: = 0.14;

8 7

4

6 + 3 1

3 28

Order does not count/replacements allowed: = 0.17;

8 + 4 1 165

4

(5 4 3)4 = 1 0.14;

8765 7

4 3 4 2 4 4

1 5 + 2 5 + 3 5 + 4

Order counts/replacements allowed:

=

671

0.16.

84 4, 096

EXAMPLE 6 What is the probability that a poker hand will have exactly one pair?

A poker hand is a 5-subset of the set of 52 cards in a full deck, so there

are

52

= N = 2, 598, 960

5

different poker hands. We thus let S be a set with N elements and assign the

uniform probability measure to S. A poker hand with one pair has two cards

of the same face value and three cards whose faces are all different among

themselves and from that of the pair. We arrive at a unique poker hand with

one pair by completing the following tasks in order:

a) Choose the face value of the pair from the 13 available face values. This can

be done in (131) = 13 ways.

b) Choose two cards with the face value selected in (a). This can be done in (42)

= 6 ways.

c) Choose the three face values for the other three cards in the hand. Since

there are 12 face values to choose from, this can be done in (123) = 220

ways.

d) Choose one card (from the four at hand) of each face value chosen in (c).

This can be done in 4 4 4 = 43 = 64 ways.

38 2 Some Probabilistic Concepts and Results

one pair. Hence, by assuming the uniform probability measure, the required

probability is equal to

1, 098, 240

0.42.

2, 598, 960

THEOREM 9 i) The number of ways in which n distinct balls can be distributed into k

distinct cells is kn.

ii) The number of ways that n distinct balls can be distributed into k distinct

cells so that the jth cell contains nj balls (nj 0, j = 1, . . . , k, kj=1 nj = n)

is

n! n

= .

n1! n2 ! nk ! n1 , n2 , , nk

iii) The number of ways that n indistinguishable balls can be distributed into

k distinct cells is

k + n 1

.

n

n 1

.

k 1

PROOF

i) Obvious, since there are k places to put each of the n balls.

ii) This problem is equivalent to partitioning the n balls into k groups, where

the jth group contains exactly nj balls with nj as above. This can be done in

the following number of ways:

n n n1 n n1 nk1 n!

= n !n ! n !.

n1 n2 nk 1 2 k

iii) We represent the k cells by the k spaces between k + 1 vertical bars and the

n balls by n stars. By fixing the two extreme bars, we are left with k + n

1 bars and stars which we may consider as k + n 1 spaces to be filled in

by a bar or a star. Then the problem is that of selecting n spaces for the n

( )

stars which can be done in k+nn1 ways. As for the second part, we now

have the condition that there should not be two adjacent bars. The n stars

k1 ( )

create n 1 spaces and by selecting k 1 of them in n1 ways to place the

2.4 Combinatorial Results 39

REMARK 5

i) The numbers nj, j = 1, . . . , k in the second part of the theorem are called

occupancy numbers.

ii) The answer to (ii) is also the answer to the following different question:

Consider n numbered balls such that nj are identical among themselves and

distinct from all others, nj 0, j = 1, . . . , k, kj=1 nj = n. Then the number of

different permutations is

n

.

n1 , n2 , , nk

Now consider the following examples for the purpose of illustrating the

theorem.

EXAMPLE 7 Find the probability that, in dealing a bridge hand, each player receives one

ace.

The number of possible bridge hands is

52 52!

N = = .

13, 13, 13, 13 ( )

4

13!

Our sample space S is a set with N elements and assign the uniform probability

measure. Next, the number of sample points for which each player, North,

South, East and West, has one ace can be found as follows:

a) Deal the four aces, one to each player. This can be done in

4 4!

= 1! 1! 1! 1! = 4! ways.

1, 1, 1, 1

b) Deal the remaining 48 cards, 12 to each player. This can be done in

48 48!

= ways.

12, 12, 12, 12 ( )

4

12!

4!48!(13!)4/[(12!)4(52!)]. Furthermore, it can be seen that this probability lies

between 0.10 and 0.11.

EXAMPLE 8 The eleven letters of the word MISSISSIPPI are scrambled and then arranged

in some order.

i) What is the probability that the four Is are consecutive letters in the

resulting arrangement?

There are eight possible positions for the first I and the remaining

probability is

( )

seven letters can be arranged in 7 distinct ways. Thus the required

1, 4 , 2

40 2 Some Probabilistic Concepts and Results

7

8

1, 4, 2 4

= 0.02.

11 165

1, 4, 4, 2

ii) What is the conditional probability that the four Is are consecutive (event

A), given B, where B is the event that the arrangement starts with M and

ends with S?

Since there are only six positions for the first I, we clearly have

5

6

( )

P AB =

2

9

=

1

21

0.05.

4, 3, 2

iii) What is the conditional probability of A, as defined above, given C, where

C is the event that the arrangement ends with four consecutive Ss?

Since there are only four positions for the first I, it is clear that

3

4

( )

P AC =

2

7

=

4

35

0.11.

1, 2, 4

Exercises

2.4.1 A combination lock can be unlocked by switching it to the left and

stopping at digit a, then switching it to the right and stopping at digit b and,

finally, switching it to the left and stopping at digit c. If the distinct digits a, b

and c are chosen from among the numbers 0, 1, . . . , 9, what is the number of

possible combinations?

2.4.2 How many distinct groups of n symbols in a row can be formed, if each

symbol is either a dot or a dash?

2.4.3 How many different three-digit numbers can be formed by using the

numbers 0, 1, . . . , 9?

2.4.4 Telephone numbers consist of seven digits, three of which are grouped

together, and the remaining four are also grouped together. How many num-

bers can be formed if:

i) No restrictions are imposed?

ii) If the first three numbers are required to be 752?

2.4 Combinatorial Results

Exercises 41

2.4.5 A certain state uses five symbols for automobile license plates such that

the first two are letters and the last three numbers. How many license plates

can be made, if:

i) All letters and numbers may be used?

ii) No two letters may be the same?

2.4.6 Suppose that the letters C, E, F, F, I and O are written on six chips and

placed into an urn. Then the six chips are mixed and drawn one by one without

replacement. What is the probability that the word OFFICE is formed?

2.4.7 The 24 volumes of the Encyclopaedia Britannica are arranged on a

shelf. What is the probability that:

i) All 24 volumes appear in ascending order?

ii) All 24 volumes appear in ascending order, given that volumes 14 and 15

appeared in ascending order and that volumes 113 precede volume 14?

2.4.8 If n countries exchange ambassadors, how many ambassadors are

involved?

2.4.9 From among n eligible draftees, m men are to be drafted so that all

possible combinations are equally likely to be chosen. What is the probability

that a specified man is not drafted?

2.4.10 Show that

n + 1

m + 1 n+1

= .

n m+1

m

2.4.11 Consider five line segments of length 1, 3, 5, 7 and 9 and choose three

of them at random. What is the probability that a triangle can be formed by

using these three chosen line segments?

2.4.12 From 10 positive and 6 negative numbers, 3 numbers are chosen at

random and without repetitions. What is the probability that their product is

a negative number?

2.4.13 In how many ways can a committee of 2n + 1 people be seated along

one side of a table, if the chairman must sit in the middle?

2.4.14 Each of the 2n members of a committee flips a fair coin in deciding

whether or not to attend a meeting of the committee; a committee member

attends the meeting if an H appears. What is the probability that a majority

will show up in the meeting?

2.4.15 If the probability that a coin falls H is p (0 < p < 1), what is the

probability that two people obtain the same number of Hs, if each one of

them tosses the coin independently n times?

42 2 Some Probabilistic Concepts and Results

2.4.16

i) Six fair dice are tossed once. What is the probability that all six faces

appear?

ii) Seven fair dice are tossed once. What is the probability that every face

appears at least once?

2.4.17 A shipment of 2,000 light bulbs contains 200 defective items and 1,800

good items. Five hundred bulbs are chosen at random, are tested and the

entire shipment is rejected if more than 25 bulbs from among those tested are

found to be defective. What is the probability that the shipment will be

accepted?

2.4.18 Show that

M M 1 M 1

= + ,

m m m 1

where N, m are positive integers and m < M.

2.4.19 Show that

r

m n m + n

x r x = r

,

x =0

where

k

= 0 if x > k.

x

2.4.20 Show that

n

n n

n

(1) j = 0.

j

i) j = 2 n ; ii )

j=0 j=0

there are supplied 5 different answers (of which only one is correct). The

student is required to answer correctly at least 25 questions in order to pass the

test. If he knows the right answers to the first 20 questions and chooses an

answer to the remaining questions at random and independently of each other,

what is the probability that he will pass the test?

2.4.22 A student committee of 12 people is to be formed from among 100

freshmen (60 male + 40 female), 80 sophomores (50 male + 30 female), 70

juniors (46 male + 24 female), and 40 seniors (28 male + 12 female). Find the

total number of different committees which can be formed under each one of

the following requirements:

i) No restrictions are imposed on the formation of the committee;

ii) Seven students are male and five female;

2.4 Combinatorial Results

Exercises 43

iii) The committee contains the same number of students from each class;

iv) The committee contains two male students and one female student from

each class;

v) The committee chairman is required to be a senior;

vi) The committee chairman is required to be both a senior and male;

vii) The chairman, the secretary and the treasurer of the committee are all

required to belong to different classes.

2.4.23 Refer to Exercise 2.4.22 and suppose that the committee is formed by

choosing its members at random. Compute the probability that the committee

to be chosen satisfies each one of the requirements (i)(vii).

2.4.24 A fair die is rolled independently until all faces appear at least once.

What is the probability that this happens on the 20th throw?

dom into the 20 envelopes. What is the probability that:

i) All 20 letters go into the right envelopes?

ii) Exactly 19 letters go into the right envelopes?

iii) Exactly 17 letters go into the right envelopes?

2.4.26 Suppose that each one of the 365 days of a year is equally likely to be

the birthday of each one of a given group of 73 people. What is the probability

that:

i) Forty people have the same birthdays and the other 33 also have the same

birthday (which is different from that of the previous group)?

ii) If a year is divided into five 73-day specified intervals, what is the probabil-

ity that the birthday of: 17 people falls into the first such interval, 23 into

the second, 15 into the third, 10 into the fourth and 8 into the fifth interval?

2.4.27 Suppose that each one of n sticks is broken into one long and one

short part. Two parts are chosen at random. What is the probability that:

i) One part is long and one is short?

ii) Both parts are either long or short?

The 2n parts are arranged at random into n pairs from which new sticks are

formed. Find the probability that:

iii) The parts are joined in the original order;

iv) All long parts are paired with short parts.

2.4.28 Derive the third part of Theorem 9 from Theorem 8(ii).

2.4.29 Three cards are drawn at random and with replacement from a stan-

dard deck of 52 playing cards. Compute the probabilities P(Aj), j = 1, . . . , 5,

where the events Aj, j = 1, . . . , 5 are defined as follows:

44 2 Some Probabilistic Concepts and Results

{

A1 = s S ; all 3 cards in s are black , }

A2 = {s S ; at least 2 cards in s are red},

A3 = {s S ; exactly 1 card in s is an ace},

A4 = {s S ; the first card in s is a diamond,

}

the second is a heart and the third is a club ,

{ }

A5 = s S ; 1 card in s is a diamond, 1 is a heart and 1 is a club .

j = 1, . . . , 5 when the cards are drawn at random but without replacement.

2.4.31 Consider hands of 5 cards from a standard deck of 52 playing

cards. Find the number of all 5-card hands which satisfy one of the following

requirements:

i) Exactly three cards are of one color;

ii) Three cards are of three suits and the other two of the remaining suit;

iii) At least two of the cards are aces;

iv) Two cards are aces, one is a king, one is a queen and one is a jack;

v) All five cards are of the same suit.

2.4.32 An urn contains nR red balls, nB black balls and nW white balls. r balls

are chosen at random and with replacement. Find the probability that:

i) All r balls are red;

ii) At least one ball is red;

iii) r1 balls are red, r2 balls are black and r3 balls are white (r1 + r2 + r3 = r);

iv) There are balls of all three colors.

2.4.33 Refer to Exercise 2.4.32 and discuss the questions (i)(iii) for r = 3 and

r1 = r2 = r3 (= 1), if the balls are drawn at random but without replacement.

2.4.34 Suppose that all 13-card hands are equally likely when a standard

deck of 52 playing cards is dealt to 4 people. Compute the probabilities P(Aj),

j = 1, . . . , 8, where the events Aj, j = 1, . . . , 8 are defined as follows:

{

A1 = s S ; s consists of 1 color cards , }

A2 = {s S ; s consists only of diamonds},

A3 = {s S ; s consists of 5 diamonds, 3 hearts, 2 clubs and 3 spades},

A4 = {s S ; s consists of cards of exactly 2 suits},

A5 = {s S ; s contains at least 2 aces},

A6 = {s S ; s does not contain aces, tens and jacks},

2.5 Product

2.4 Combinatorial

Probability Results

Spaces 45

{

A7 = s S ; s consists of 3 aces, 2 kings and exactly 7 red cards , }

A8 = {s S ; s consists of cards of all different denominations}.

and also A as follows:

{ }

A j = s S ; s contains exactly j tens ,

A = {s S ; s contains exactly 7 red cards}.

compare the numbers P(Aj), P(Aj|A).

2.4.36 Let S be the set of all n3 3-letter words of a language and let P be the

equally likely probability function on the events of S. Define the events A, B

and C as follows:

{

A = s S ; s begins with a specific letter , }

{ (

B = s S ; s has the specified letter mentioned in the definition of A )

in the middle entry , }

{

C = s S ; s has exactly two of its letters the same . }

Then show that:

i) P(A B) = P(A)P(B);

ii) P(A C) = P(A)P(C);

iii) P(B C) = P(B)P(C);

iv) P(A B C) P(A)P(B)P(C).

Thus the events A, B, C are pairwise independent but not mutually

independent.

The concepts discussed in Section 2.3 can be stated precisely by utilizing more

technical language. Thus, if we consider the experiments E1 and E2 with re-

spective probability spaces (S1, A1, P1) and (S2, A2, P2), then the compound

experiment (E1, E2) = E1 E2 has sample space S = S1 S2 as defined earlier.

The appropriate -field A of events in S is defined as follows: First define the

class C by:

{

C = A1 A2 ; A1 A1 , A2 A2 , }

where A1 A2 = {(s , s ); s A , s

1 2 1 1 2 }

A2 .

46 2 Some Probabilistic Concepts and Results

1). Next, define on C the set function P by P(A1 A2) = P1(A1)P2(A2). It can be

shown that P determines uniquely a probability measure on A (by means of

the so-called Carathodory extension theorem). This probability measure is

usually denoted by P1 P2 and is called the product probability measure (with

factors P1 and P2), and the probability space (S, A, P) is called the product

probability space (with factors (Sj, Aj, Pj), j = 1, 2). It is to be noted that events

which refer to E1 alone are of the form B1 = A1 S2, A1 A1, and those

referring to E2 alone are of the form B2 = S1 A2, A2 A2. The experiments E1

and E2 are then said to be independent if P(B1 B2) = P(B1)P(B2) for all events

B1 and B2 as defined above.

For n experiments Ej, j = 1, 2, . . . , n with corresponding probability spaces

(Sj, Aj, Pj), the compound experiment (E1, . . . , En) = E1 En has prob-

ability space (S, A, P), where

S = S1 S n = {(s , , s ); s S , j = 1, 2, , n},

1 n j j

{

C = A1 An ; A j A j , j = 1, 2, , n , }

and P is the unique probability measure defined on A through the

relationships

( ) ( ) ( )

P A1 An = P A1 P An , A j A j , j = 1, 2, , n.

The probability measure P is usually denoted by P1 Pn and is called the

product probability measure (with factors Pj, j = 1, 2, . . . , n), and the probabil-

ity space (S, A, P) is called the product probability space (with factors (Sj, Aj,

Pj), j = 1, 2, . . . , n). Then the experiments Ej, j = 1, 2, . . . , n are said to be

independent if P(B1 B2) = P(B1) P(B2), where Bj is defined by

B j = S1 S j 1 A j S j + 1 S n , j = 1, 2, , n.

Let A1, A2 be two sub--fields of A. We say that A1, A2 are independent if

P(A1 A2) = P(A1)P(A2) for any A1 A1, A2 A2. More generally, the

-fields Aj, j = 1, 2, . . . , n (sub--fields of A) are said to be independent if

n n

P I A j = P A j

j =1 j =1

( ) for any A j A j , j = 1, 2, , n.

At this point, notice that the factor -fields Aj, j = 1, 2, . . . , n may be

considered as sub--fields of the product -field A by identifying Aj with Bj,

where the Bjs are defined above. Then independence of the experiments Ej,

j = 1, 2, . . . , n amounts to independence of the corresponding -fields Aj,

j = 1, 2, . . . , n (looked upon as sub--fields of the product -field A).

2.6* The

2.4Probability

Combinatorial

of Matchings

Results 47

Exercises

2.5.1 Form the Cartesian products A B, A C, B C, A B C, where

A = {stop, go}, B = {good, defective), C = {(1, 1), (1, 2), (2, 2)}.

i) (A B)c = (A Bc) + (Ac B) + (Ac Bc);

ii) (A B) (C D) = (A C) (B D);

iii) (A B) (C D) = (A C) (B D) [(A Cc) (Bc D)

+ (Ac C) (B Dc )].

In this section, an important result, Theorem 10, is established providing an

expression for the probability of occurrence of exactly m events out of possible

M events. The theorem is then illustrated by means of two interesting exam-

ples. For this purpose, some additional notation is needed which we proceed to

introduce. Consider M events Aj, j = 1, 2, . . . , M and set

S0 = 1,

M

S1 = P A j ,

j =1

( )

S2 =

1 j1 < j2 M

(

P A j1 A j2 , )

M

Sr =

1 j1 < j2 < < jr M

( )

P A j1 A j2 A jr ,

M

(

S M = P A1 A2 AM . )

Let also

Bm = exactly

C m = at least m of the events A j , j = 1, 2, , M occur.

Dm = at most

Then we have

48 2 Some Probabilistic Concepts and Results

m + 1 m + 2 M

( ) ( )

M m

P Bm = S m S m+1 + S m+ 2 + 1 SM (2)

m m m

which for m = 0 is

( ) ( )

M

P B0 = S0 S1 + S 2 + 1 SM , (3)

and

( ) ( ) ( )

P C m = P Bm + P Bm+1 + + P BM , ( ) (4)

and

( ) ( ) ( )

P Dm = P B0 + P B1 + + P Bm . ( ) (5)

For the proof of this theorem, all that one has to establish is (2), since (4)

and (5) follow from it. This will be done in Section 5.6 of Chapter 5. For a proof

where S is discrete the reader is referred to the book An Introduction to

Probability Theory and Its Applications, Vol. I, 3rd ed., 1968, by W. Feller, pp.

99100.

The following examples illustrate the above theorem.

EXAMPLE 9 The matching problem (case of sampling without replacement). Suppose that

we have M urns, numbered 1 to M. Let M balls numbered 1 to M be inserted

randomly in the urns, with one ball in each urn. If a ball is placed into the urn

bearing the same number as the ball, a match is said to have occurred.

i) Show the probability of at least one match is

1 1

( ) 1

M +1

1 + + 1 1 e 1 0.63

2! 3! M!

for large M, and

ii) exactly m matches will occur, for m = 0, 1, 2, . . . , M is

1

1 1

( ) 1

M m

1 1 + + + 1

m! 2! 3! ( )

M m !

1 M m

( ) 1 1 1

k

= 1

m! k = 0

k! m!

e for M m large.

DISCUSSION To describe the distribution of the balls among the urns, write

an M-tuple (z1, z2, . . . , zM) whose jth component represents the number of the

ball inserted in the jth urn. For k = 1, 2, . . . , M, the event Ak that a match will

occur in the kth urn may be written Ak = {(z1, . . . , zM) ! M; zj integer, 1 zj

2.6* The

2.4Probability of Matchings

Combinatorial Results 49

r unequal integers k1, k2, . . . , kr,, from 1 to M,

) ( M! ) .

M r !

(

P Ak1 Ak2 Akr =

( )

M M r ! 1

Sr = = .

r M! r!

This implies the desired results.

EXAMPLE 10 Coupon collecting (case of sampling with replacement). Suppose that a manu-

facturer gives away in packages of his product certain items (which we take to

be coupons), each bearing one of the integers 1 to M, in such a way that each

of the M items is equally likely to be found in any package purchased. If n

packages are bought, show that the probability that exactly m of the integers,

1 to M, will not be obtained is equal to

n

M M m k M m m + k

( )

1

m k = 0 k

1 M .

Many variations and applications of the above problem are described in

the literature, one of which is the following. If n distinguishable balls are

distributed among M urns, numbered 1 to M, what is the probability that there

will be exactly m urns in which no ball was placed (that is, exactly m urns

remain empty after the n balls have been distributed)?

DISCUSSION To describe the coupons found in the n packages purchased,

we write an n-tuple (z1, z2, , zn), whose jth component zj represents the

number of the coupon found in the jth package purchased. We now define the

events A1, A2, . . . , AM. For k = 1, 2, . . . , M, Ak is the event that the number k

will not appear in the sample, that is,

( )

Ak = z1 , , zn ! n ; z j integer, 1 z j M , z j k, j = 1, 2, , n.

It is easy to see that we have the following results:

n n

M 1 1

( )

P Ak =

M

= 1 ,

M

k = 1, 2, , M ,

n n

M 2 2 k1 = 1, 2, , n

(

P Ak Ak1 2

) =

M

= 1 ,

M k2 = k1 + 1, , n

and, in general,

50 2 Some Probabilistic Concepts and Results

k1 = 1, 2, , n

n

r k2 = k1 + 1, , n

( 1 2 r

M

)

P Ak Ak Ak = 1 ,

M

kr = kr 1 + 1, , n.

n

M r

S r = 1 , r = 0, 1, , M . (6)

r M

Let Bm be the event that exactly m of the integers 1 to M will not be found in

the sample. Clearly, Bm is the event that exactly m of the events A1, . . . , AM

will occur. By relations (2) and (6), we have

n

M

r M r

( ) ( )

r m

P Bm = 1 1 M

r =m m r

n

M M m k M m m + k

= 1

m k=0 k

( )

1 M ,

(7)

m + k M M M m

= . (8)

m m + k m k

This section is concluded with the following important result stated as a

theorem.

THEOREM 11 Let A and B be two disjoint events. Then in a series of independent trials, show

that:

( )

P A

(

P A occurs before B occurs = ) .

( ) ( )

P A +P B

Then, clearly, required the event is the sum of the events

A1c B1c Anc Bnc An+1 ,

and therefore

2.6* The

2.4Probability

Combinatorial

of Matchings

Results 51

(

P A occurs before B occurs )

[ (

= P A1 + A1c B1c A2 + A1c B1c A2c B2c A3 ) ( )

(

+ + A1c B1c Anc Bnc An+1 + ) ]

( ) (

= P A1 + P A1c B1c A2 + P A1c B1c A2c B2c A3 ) ( )

( )

+ + P A1c B1c Anc Bnc An+1 +

= P( A ) + P( A B ) P( A ) + P( A B ) P( A B ) P( A )

1 1

c

1

c

2 1

c

1

c c

2

c

2 3

1

c

1

c c

n

c

n n+1

= P( A) + P( A B )P( A) + P ( A B )P( A)

c c 2 c c

+ + P ( A B )P( A)

n c c

= P( A)[1 + P( A B ) + P ( A B ) + + P ( A B ) + ]

c c 2 c c n c c

( ) 1 P A1 B

=P A .

( c c

)

But

( )

P Ac Bc = P A B = 1 P A B

c

( ) ( )

= 1 P A+ B = 1 P A P B , ( ) ( ) ( )

so that

(

1 P Ac Bc = P A + P B . ) ( ) ()

Therefore

( )

P A

(

P A occurs before B occurs = ) ,

( ) ( )

P A +P B

as asserted.

It is possible to interpret B as a catastrophic event, and A as an event

consisting of taking certain precautionary and protective actions upon the

energizing of a signaling device. Then the significance of the above probability

becomes apparent. As a concrete illustration, consider the following simple

example (see also Exercise 2.6.3).

52 2 Some Probabilistic Concepts and Results

playing cards, calculate the probability that an ace occurs before a picture.

Then P(A) = 4

52

= 1

13

and P(B) = 12

52

= 4

13

, so that P(A occurs before B occurs)

= 131 134 = 14 .

Exercises

2.6.1 Show that

m + k M M M m

= ,

m m + k m k

as asserted in relation (8).

2.6.2 Verify the transition in (7) and that the resulting expression is indeed

the desired result.

2.6.3 Consider the following game of chance. Two fair dice are rolled repeat-

edly and independently. If the sum of the outcomes is either 7 or 11, the player

wins immediately, while if the sum is either 2 or 3 or 12, the player loses

immediately. If the sum is either 4 or 5 or 6 or 8 or 9 or 10, the player continues

rolling the dice until either the same sum appears before a sum of 7 appears in

which case he wins, or until a sum of 7 appears before the original sum appears

in which case the player loses. It is assumed that the game terminates the first

time the player wins or loses. What is the probability of winning?

3.1 Soem General Concepts 53

Chapter 3

Their Distributions

Given a probability space (S, class of events, P), the main objective of prob-

ability theory is that of calculating probabilities of events which may be of

importance to us. Such calculations are facilitated by a transformation of the

sample space S, which may be quite an abstract set, into a subset of the real

line ! with which we are already familiar. This is, actually, achieved by the

introduction of the concept of a random variable. A random variable (r.v.) is

a function (in the usual sense of the word), which assigns to each sample point

s S a real number, the value of the r.v. at s. We require that an r.v. be a well-

behaving function. This is satisfied by stipulating that r.v.s are measurable

functions. For the precise definition of this concept and related results, the

interested reader is referred to Section 3.5 below. Most functions as just

defined, which occur in practice are, indeed, r.v.s, and we leave the matter to

rest here. The notation X(S) will be used for the set of values of the r.v. X, the

range of X.

Random variables are denoted by the last letters of the alphabet X, Y, Z,

etc., with or without subscripts. For a subset B of !, we usually denote by

(X B) the following event in S: (X B) = {s S; X(s) B} for simplicity. In

particular, (X = x) = {s S; X(s) = x}. The probability distribution function (or

just the distribution) of an r.v. X is usually denoted by PX and is a probability

function defined on subsets of ! as follows: PX (B) = P(X B). An r.v. X is said

to be of the discrete type (or just discrete) if there are countable (that is, finitely

many or denumerably infinite) many points in !, x1, x2, . . . , such that PX({xj})

> 0, j 1, and j PX({xj})(= j P(X = xj)) = 1. Then the function fX defined on the

entire ! by the relationships:

( ) ({ })(= P(X = x ))

fX x j = PX x j j for x = x j ,

53

54 3 On Random Variables and Their Distributions

()

fX x 0 for all x, and f (x ) = 1.

j

X j

( ) f (x ).

P X B =

x j B

X j

X(s) B}, all we have to do is to sum up the values of fX(xj) for all those xjs

which lie in B; this assumes, of course, that the function fX is known. The

function fX is called the probability density function (p.d.f.) of X. The distribu-

tion of a discrete r.v. will also be referred to as a discrete distribution. In the

following section, we will see some discrete r.v.s (distributions) often occur-

ring in practice. They are the Binomial, Poisson, Hypergeometric, Negative

Binomial, and the (discrete) Uniform distributions.

Next, suppose that X is an r.v. which takes values in a (finite or infinite but

proper) interval I in ! with the following qualification: P(X = x) = 0 for every

single x in I. Such an r.v. is called an r.v. of the continuous type (or just a

continuous r.v.). Also, it often happens for such an r.v. to have a function fX

satisfying the properties fX(x) 0 for all x I, and P(X J) = J fX(x)dx for any

sub-interval J of I. Such a function is called the probability density function

(p.d.f.) of X in analogy with the discrete case. It is to be noted, however, that

here fX(x) does not represent the P(X = x)! A continuous r.v. X with a p.d.f. fX

is called absolutely continuous to differentiate it from those continuous r.v.s

which do not have a p.d.f. In this book, however, we are not going to concern

ourselves with non-absolutely continuous r.v.s. Accordingly, the term con-

tinuous r.v. will be used instead of absolutely continuous r.v. Thus, the r.v.s

to be considered will be either discrete or continuous (= absolutely continu-

ous). Roughly speaking, the idea that P(X = x) = 0 for all x for a continuous r.v.

may be interpreted that X takes on too many values for each one of them to

occur with positive probability. The fact that P(X = x) also follows formally by

the fact that P(X = x) = xx fX(y)dy, and this is 0. Other interpretations are also

possible. It is true, nevertheless, that X takes values in as small a neighborhood

of x as we please with positive probability. The distribution of a continuous r.v.

is also referred to as a continuous distribution. In Section 3.3, we will discuss

some continuous r.v.s (distributions) which occur often in practice. They are

the Normal, Gamma, Chi-square, Negative Exponential, Uniform, Beta,

Cauchy, and Lognormal distributions. Reference will also be made to t and F

r.v.s (distributions).

Often one is given a function f and is asked whether f is a p.d.f. (of some

r.v.). All one has to do is to check whether f is non-negative for all values of its

argument, and whether the sum or integral of its values (over the appropriate

set) is equal to 1.

3.2 Discrete Random Variables

3.1 Soem

(and General

RandomConcepts

Vectors) 55

takes values in the plane or the three-dimensional space or, more generally, in

the k-dimensional space ! k, it is called a k-dimensional random vector (r.

vector) and is denoted by X. Thus, an r.v. is a one-dimensional r. vector. The

distribution of X, PX, is defined as in the one-dimensional case by simply

replacing B with subsets of ! k. The r. vector X is discrete if P(X = xj) > 0, j =

1, 2, . . . with j P(X = xj) = 1, and the function fX(x) = P(X = xj) for x = xj, and

fX(x) = 0 otherwise is the p.d.f. of X. Once again, P(X B) = xjB fX(xj) for B

subsets of !k. The r. vector X is (absolutely) continuous if P(X = x) = 0 for all

x I, but there is a function fX defined on !k such that:

()

fX x 0 (

for all x ! k , and P X J = fX x dx ) J

()

for any sub-rectangle J of I. The function fX is the p.d.f. of X. The distribution

of a k-dimensional r. vector is also referred to as a k-dimensional discrete or

(absolutely) continuous distribution, respectively, for a discrete or (abso-

lutely) continuous r. vector. In Sections 3.2 and 3.3, we will discuss two repre-

sentative multidimensional distributions; namely, the Multinomial (discrete)

distribution, and the (continuous) Bivariate Normal distribution.

We will write f rather than fX when no confusion is possible. Again, when

one is presented with a function f and is asked whether f is a p.d.f. (of some r.

vector), all one has to check is non-negativity of f, and that the sum of its values

or its integral (over the appropriate space) is equal to 1.

3.2.1 Binomial

The Binomial distribution is associated with a Binomial experiment; that is, an

experiment which results in two possible outcomes, one usually termed as a

success, S, and the other called a failure, F. The respective probabilities

are p and q. It is to be noted, however, that the experiment does not really

have to result in two outcomes only. Once some of the possible outcomes are

called a failure, any experiment can be reduced to a Binomial experiment.

Here, if X is the r.v. denoting the number of successes in n binomial experi-

ments, then

n

( ) {

X S = 0, 1, 2, , n , } ( ) ()

P X = x = f x = p x q n x ,

x

where 0 < p < 1, q = 1 p, and x = 0, 1, 2, . . . , n. That this is in fact a p.d.f.

follows from the fact that f(x) 0 and

n n

n

f ( x ) = x p x q n x = ( p + q)

n

= 1n = 1.

x =0 x =0

56 3 On Random Variables and Their Distributions

{ }

S = S, F S, F{ } (n copies).

In particular, for n = 1, we have the Bernoulli or Point Binomial r.v. The r.v.

X may be interpreted as representing the number of Ss (successes) in

the compound experiment E E (n copies), where E is the experiment

resulting in the sample space {S, F} and the n experiments are independent

(or, as we say, the n trials are independent). f(x) is the probability that exactly

x Ss occur. In fact, f(x) = P(X = x) = P(of all n sequences of Ss and Fs

with exactly x Ss). The probability of one such a sequence is pxqnx by the

independence of the trials and this also does not depend on the particular

sequence we are considering. Since there are ( xn) such sequences, the result

follows.

The distribution of X is called the Binomial distribution and the quantities

n and p are called the parameters of the Binomial distribution. We denote the

Binomial distribution by B(n, p). Often the notation X B(n, p) will be used

to denote the fact that the r.v. X is distributed as B(n, p). Graphs of the p.d.f.

of the B(n, p) distribution for selected values of n and p are given in Figs. 3.1

and 3.2.

f (x)

0.25

n ! 12

0.20 1

p! 4

0.15

0.10

0.05

x

0 1 2 3 4 5 6 7 8 9 10 11 12 13

Figure 3.1 Graph of the p.d.f. of the Binomial distribution for n = 12, p = 14 .

()

f 0 = 0.0317 ()

f 7 = 0.0115

f (1) = 0.1267 f ( 8 ) = 0.0024

f (2 ) = 0.2323 f (9) = 0.0004

f ( 3) = 0.2581 f (10 ) = 0.0000

f ( 4 ) = 0.1936 f (11) = 0.0000

f ( 5) = 0.1032 f (12 ) = 0.0000

f (6) = 0.0401

3.2 3.1 Soem

Discrete Random Variables (and General

RandomConcepts

Vectors) 57

f (x)

0.25

0.20 n ! 10

1

0.15 p! 2

0.10

0.05

x

0 1 2 3 4 5 6 7 8 9 10

Figure 3.2 Graph of the p.d.f. of the Binomial distribution for n = 10, p = 12 .

()

f 0 = 0.0010 ()

f 6 = 0.2051

f (1) = 0.0097 f ( 7) = 0.1172

f (2 ) = 0.0440 f ( 8 ) = 0.0440

f ( 3) = 0.1172 f (9) = 0.0097

f ( 4 ) = 0.2051 f (10 ) = 0.0010

f ( 5) = 0.2460

3.2.2 Poisson

x

( ) {

X S = 0, 1, 2, , } ( ) ()

P X = x = f x = e

x!

,

x

f ( x ) = e x! = e

e = 1.

x =0 x =0

P(). is called the parameter of the distribution. Often the notation X

P() will be used to denote the fact that the r.v. X is distributed as P().

The Poisson distribution is appropriate for predicting the number of

phone calls arriving at a given telephone exchange within a certain period

of time, the number of particles emitted by a radioactive source within a

certain period of time, etc. The reader who is interested in the applications

of the Poisson distribution should see W. Feller, An Introduction to

Probability Theory, Vol. I, 3rd ed., 1968, Chapter 6, pages 156164, for further

examples.

In Theorem 1 in Section 3.4, it is shown that the Poisson distribution

58 3 On Random Variables and Their Distributions

pose that X B(n, p), xwhere n is large and p is small. Then P X = x = ( )

( xn ) px (1 p) e np ( npx !) , x 0 . For the graph of the p.d.f. of the P()

n x

A visualization of such an approximation may be conceived by stipulating

that certain events occur in a time interval [0,t] in the following manner: events

occurring in nonoverlapping subintervals are independent; the probability

that one event occurs in a small interval is approximately proportional to its

length; and two or more events occur in such an interval with probability

approximately 0. Then dividing [0,t] into a large number n of small intervals of

length t/n, we have that the probability that exactly x events occur in [0,t] is

approximately ( nx )( nt ) (1 nt ) , where is the factor of proportionality.

x n x

( nx )( nt ) (1 nt ) e t (xt )! . Thus Binomial probabilities are approximated by

x n x

Poisson probabilities.

f (x)

0.20

0.15

0.10

0.05

x

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15

Figure 3.3 Graph of the p.d.f. of the Poisson distribution with = 5.

()

f 0 = 0.0067 ()

f 9 = 0.0363

f (1) = 0.0337 f (10 ) = 0.0181

f (2 ) = 0.0843 f (11) = 0.0082

f ( 3) = 0.1403 f (12 ) = 0.0035

f ( 4 ) = 0.1755 f (13) = 0.0013

f ( 5) = 0.1755 f (14 ) = 0.0005

f (6) = 0.1462 f (15) = 0.0001

f ( 7) = 0.1044

f ( 8 ) = 0.0653 ()

f n is negligible for n 16.

3.2 Discrete Random Variables

3.1 Soem

(and General

RandomConcepts

Vectors) 59

3.2.3 Hypergeometric

m n

x r x

( ) {

X S = 0, 1, 2, , r , } ()

f x =

m + n

,

r

r r

m n m + n

f ( x ) = m + n x r x = m + n

1 1

= 1.

x =0 x =0 r

r r

The distribution of X is called the Hypergeometric distribution and arises in

situations like the following. From an urn containing m red balls and n black

balls, r balls are drawn at random without replacement. Then X represents the

number of red balls among the r balls selected, and f(x) is the probability that

this number is exactly x. Here S = {all r-sequences of Rs and Bs}, where R

stands for a red ball and B stands for a black ball. The urn/balls model just

described is a generic model for situations often occurring in practice. For

instance, the urn and the balls may be replaced by a box containing certain

items manufactured by a certain process over a specified period of time, out of

which m are defective and n meet set specifications.

r + x 1 x

( ) {

X S = 0, 1, 2, , } ()

f x = pr

x

q ,

r + x 1 x pr pr

f ( x ) = p

r

q = = = 1.

x ( )

r

x =0 x =0 1 q pr

1

n + j 1 j

= x , x < 1.

(1 x) j=0

n

j

bution occurs in situations which have as a model the following. A Binomial

experiment E, with sample space {S, F}, is repeated independently until exactly

r Ss appear and then it is terminated. Then the r.v. X represents the number

of times beyond r that the experiment is required to be carried out, and f(x) is

the probability that this number of times is equal to x. In fact, here S =

60 3 On Random Variables and Their Distributions

{all (r + x)-sequences of Ss and Fs such that the rth S is at the end of the

sequence}, x = 0, 1, . . . and f(x) = P(X = x) = P[all (r + x)-sequences as above

for a specified x]. The probability of one such sequence is pr1qxp by the

independence assumption, and hence

r + x 1 r 1 x r r + x 1 x

()

f x =

x

p q p= p

x

q .

The above interpretation also justifies the name of the distribution. For r = 1,

we get the Geometric (or Pascal) distribution, namely f(x) = pqx, x = 0, 1, 2, . . . .

( ) {

X S = 0, 1, , n 1 , } ()

f x =

1

n

, x = 0, 1, , n 1.

f (x)

n!5

1 Figure 3.4 Graph of the p.d.f. of a Discrete

5 Uniform distribution.

x

0 1 2 3 4

3.2.6 Multinomial

Here

k

() ( )

X S = x = x1 , , xk ; x j 0, j = 1, 2, , k, x j = n,

j =1

k

()

f x =

n!

x1! x 2 ! x k !

p1x 1 p2x 2 pkx k , p j > 0, j = 1, 2, , k, p j = 1.

j =1

f ( x) = ( )

n! n

p1x1 pkxk = p1 + + pk = 1n = 1,

X x 1 , , xk x1! xk !

where the summation extends over all xjs such that xj 0, j = 1, 2, . . . , k,

kj=1 xj = n. The distribution of X is also called the Multinomial distribution and

n, p1, . . . , pk are called the parameters of the distribution. This distribution

occurs in situations like the following. A Multinomial experiment E with k

possible outcomes Oj, j = 1, 2, . . . , k, and hence with sample space S = {all

n-sequences of Ojs}, is carried out n independent times. The probability of

the Ojs occurring is pj, j = 1, 2, . . . k with pj > 0 and kj=1 p j = 1 . Then X is the

random vector whose jth component Xj represents the number of times xj

the outcome Oj occurs, j = 1, 2, . . . , k. By setting x = (x1, . . . , xk), then f is the

3.1 Soem General Exercises

Concepts 61

probability that the outcome Oj occurs exactly xj times. In fact f(x) = P(X = x)

= P(all n-sequences which contain exactly xj Ojs, j = 1, 2, . . . , k). The prob-

ability of each one of these sequences is p1x1 pkxk by independence, and since

there are n!/( x1! xk ! ) such sequences, the result follows.

The fact that the r. vector X has the Multinomial distribution with param-

eters n and p1, . . . , pk may be denoted thus: X M(n; p1, . . . , pk).

REMARK 1 When the tables given in the appendices are not directly usable

because the underlying parameters are not included there, we often resort to

linear interpolation. As an illustration, suppose X B(25, 0.3) and we wish

to calculate P(X = 10). The value p = 0.3 is not included in the Binomial

Tables in Appendix III. However, 164 = 0.25 < 0.3 < 0.3125 = 165 and the

probabilities P(X = 10), for p = 164 and p = 165 are, respectively, 0.9703

and 0.8756. Therefore linear interpolation produces the value:

0.3 0.25

(

0.9703 0.9703 0.8756 ) 0.3125 0.25

= 0.8945.

Likewise for other discrete distributions. The same principle also applies

appropriately to continuous distributions.

REMARK 2 In discrete distributions, we are often faced with calculations of

the form x =1 x x . Under appropriate conditions, we may apply the following

approach:

d x d

d

x x

= x x 1 = = x

= = .

d d d 1

( )

2

x =1 x =1 x =1 x =1 1

x=2 (

x x1 ) x2

.

Exercises

3.2.1 A fair coin is tossed independently four times, and let X be the r.v.

defined on the usual sample space S for this experiment as follows:

()

X s = the number of H s in s.

iii) What is the distribution of X?

iii) What is the partition of S induced by X?

3.2.2 It has been observed that 12.5% of the applicants fail in a certain

screening test. If X stands for the number of those out of 25 applicants who fail

to pass the test, what is the probability that:

62 3 On Random Variables and Their Distributions

iii) X 1?

iii) X 20?

iii) 5 X 20?

3.2.3 A manufacturing process produces certain articles such that the prob-

ability of each article being defective is p. What is the minimum number, n, of

articles to be produced, so that at least one of them is defective with probabil-

ity at least 0.95? Take p = 0.05.

3.2.4 If the r.v. X is distributed as B(n, p) with p > 12 , the Binomial Tables

in Appendix III cannot be used directly. In such a case, show that:

ii) P(X = x) = P(Y = n x), where Y B(n, q), x = 0, 1, . . . , n, and q = 1 p;

ii) Also, for any integers a, b with 0 a < b n, one has: P(a X b) =

P(n b Y n a), where Y is as in part (i).

3.2.5 Let X be a Poisson distributed r.v. with parameter . Given that

P(X = 0) = 0.1, compute the probability that X > 5.

3.2.6 Refer to Exercise 3.2.5 and suppose that P(X = 1) = P(X = 2). What is

the probability that X < 10? If P(X = 1) = 0.1 and P(X = 2) = 0.2, calculate the

probability that X = 0.

3.2.7 It has been observed that the number of particles emitted by a radio-

active substance which reach a given portion of space during time t follows

closely the Poisson distribution with parameter . Calculate the probability

that:

iii) No particles reach the portion of space under consideration during

time t;

iii) Exactly 120 particles do so;

iii) At least 50 particles do so;

iv) Give the numerical values in (i)(iii) if = 100.

3.2.8 The phone calls arriving at a given telephone exchange within one

minute follow the Poisson distribution with parameter = 10. What is the

probability that in a given minute:

iii) No calls arrive?

iii) Exactly 10 calls arrive?

iii) At least 10 calls arrive?

3.2.9 (Truncation of a Poisson r.v.) Let the r.v. X be distributed as Poisson

with parameter and define the r.v. Y as follows:

(

Y = X if X k a given positive integer ) and Y = 0 otherwise.

Find:

3.1 Soem General Exercises

Concepts 63

ii) P(Y = 0).

3.2.10 A university dormitory system houses 1,600 students, of whom 1,200

are undergraduates and the remaining are graduate students. From the com-

bined list of their names, 25 names are chosen at random. If X stands for the

r.v. denoting the number of graduate students among the 25 chosen, what is

the probability that X 10?

3.2.11 (Multiple Hypergeometric distribution) For j = 1, . . . , k, consider an

urn containing nj balls with the number j written on them. n balls are drawn at

random and without replacement, and let Xj be the r.v. denoting the number

of balls among the n ones with the number j written on them. Then show that

the joint distribution of Xj, j = 1, . . . , k is given by

(

P X j = x j , j = 1, , k = ) (

k

( ),

nj

j =1 x j

n1 + + nk

n )

k

0 x j n j , j = 1, , k, x j = n.

j =1

3.2.12 Refer to the manufacturing process of Exercise 3.2.3 and let Y be the

r.v. denoting the minimum number of articles to be manufactured until the

first two defective articles appear.

ii) Show that the distribution of Y is given by

( ) ( )( )

y2

P Y = y = p2 y 1 1 p , y = 2, 3, ;

3.2.13 Show that the function f(x) = ( 12 )xIA(x), where A = {1, 2, . . .}, is a p.d.f.

3.2.14 For what value of c is the function f defined below a p.d.f.?

() ()

f x = c x I A x , where {

A = 0, 1, 2, } (0 < < 1).

3.2.15 Suppose that the r.v. X takes on the values 0, 1, . . . with the following

probabilities:

() (

f j =P X=j = ) c

3j

, j = 0, 1, ;

Compute the following probabilities:

iii) P(X 10);

iii) P(X A), where A = {j; j = 2k + 1, k = 0, 1, . . .};

iv) P(X B), where B = {j; j = 3k + 1, k = 0, 1, . . .}.

64 3 On Random Variables and Their Distributions

3.2.16 There are four distinct types of human blood denoted by O, A, B and

AB. Suppose that these types occur with the following frequencies: 0.45, 0.40,

0.10, 0.05, respectively. If 20 people are chosen at random, what is the prob-

ability that:

ii) All 20 people have blood of the same type?

ii) Nine people have blood type O, eight of type A, two of type B and one of

type AB?

3.2.17 A balanced die is tossed (independently) 21 times and let Xj be the

number of times the number j appears, j = 1, . . . , 6.

ii) What is the joint p.d.f. of the Xs?

ii) Compute the probability that X1 = 6, X2 = 5, X3 = 4, X4 = 3, X5 = 2, X6 = 1.

3.2.18 Suppose that three coins are tossed (independently) n times and

define the r.v.s Xj, j = 0, 1, 2, 3 as follows:

3.2.19 Let X be an r.v. distributed as P(), and set E = {0, 2, . . . } and O = {1,

3, . . . }. Then:

ii) In terms of , calculate the probabilities: P(X E) and P(X O);

ii) Find the numerical values of the probabilities in part (i) for = 5. (Hint: If

k k

SE = kE k! and SO = kO k! , notice that SE + SO = e, and SE SO =

e .)

3.2.20 The following recursive formulas may be used for calculating

Binomial, Poisson and Hypergeometric probabilities. To this effect, show

that:

iii) If X B(n, p), then f ( x + 1) = qp nx+x1 f ( x), x = 0, 1, . . . , n 1;

iii) If X P(), then f ( x + 1) = x+1 f ( x), x = 0, 1, . . . ;

iii) If X has the Hypergeometric distribution, then

(m x)(r x) f x ,

(

f x+1 = )

(n r + x + 1)(x + 1)

() { }

x = 0, 1, , min m, r .

3.2.21

i) Suppose the r.v.s X1, . . . , Xk have the Multinomial distribution, and let

j be a fixed number from the set {1, . . . , k}. Then show that Xj is distributed as

B(n, pj);

ii) If m is an integer such that 2 m k 1 and j1, . . . , jm are m distinct integers

from the set {1, . . . , k}, show that the r.v.s Xj , . . . , Xj have Multinomial

1 m

1 m

3.3 Continuous Random Variables

3.1 Soem

(and General

RandomConcepts

Vectors) 65

3.2.22 (Polyas urn scheme) Consider an urn containing b black balls and r

red balls. One ball is drawn at random, is replaced and c balls of the same color

as the one drawn are placed into the urn. Suppose that this experiment is

repeated independently n times and let X be the r.v. denoting the number of

black balls drawn. Then show that the p.d.f. of X is given by

( )(

b b + c b + 2c b + x 1 c [ ( )]

)

n r (r + c ) [r + ( n x 1)c ]

(

P X =x = ) .

x b + r b + r + c ( )( )

( )

b + r + 2c b + r + m 1 c [ ( )]

(This distribution can be used for a rough description of the spread of conta-

gious diseases. For more about this and also for a certain approximation to the

above distribution, the reader is referred to the book An Introduction to

Probability Theory and Its Applications, Vol. 1, 3rd ed., 1968, by W. Feller, pp.

120121 and p. 142.)

3.3.1 Normal (or Gaussian)

x

( )

2

( )

X S = !, f x = () 1

exp

2 2

,

x ! .

2

We say that X is distributed as normal (, 2), denoted by N(, 2), where ,

2 are called the parameters of the distribution of X which is also called

the Normal distribution ( = mean, !, 2 = variance, > 0). For = 0,

= 1, we get what is known as the Standard Normal distribution, denoted

by N(0, 1). Clearly f(x) > 0; that I = f x dx = 1 is proved by showing that ()

I = 1. In fact,

2

()

I 2 = f x dx = f x dx f y dy () ()

x 2

( y

) ( )

2

1 1

dx exp dy

2

= exp

2 2 2 2

1 1 1

2 2

= e z 2

dz e 2

d ,

2

(, ). Thus

66 3 On Random Variables and Their Distributions

f (x)

0.8

$ ! 0.5

0.6

0.4

$!1

0.2

$!2

x

"2 "1 0 1 2 3 4 5

N(#, $ 2 )

Figure 3.5 Graph of the p.d.f. of the Normal distribution with = 1.5 and several values of .

1 (

z 2 + 2 ) 2 dz d = 1 2

2

2

I2 = e e r 2

r dr d

2 0 0

1 2

r dr d = e r

2 2 2

I2 = e r 2 2

r dr = e r 2

0 = 1;

2 0 0 0

It is easily seen that f(x) is symmetric about x = , that is, f( x) =

f( + x) and that f(x) attains its maximum at x = which is equal to 1/( 2 ).

From the fact that

max f x = () 1

x !

2

and the fact that

f ( x)dx = 1,

we conclude that the larger is, the more spread-out f(x) is and vice versa. It

is also clear that f(x) 0 as x . Taking all these facts into consideration,

we are led to Fig. 3.5.

The Normal distribution is a good approximation to the distribution of

grades, heights or weights of a (large) group of individuals, lifetimes of various

manufactured items, the diameters of hail hitting the ground during a storm,

the force required to punctuate a cardboard, errors in numerous measure-

ments, etc. However, the main significance of it derives from the Central Limit

Theorem to be discussed in Chapter 8, Section 8.3.

3.3.2 Gamma

( ) (

X S = ! actually X S = 0, ( ) ( ))

3.3 Continuous Random Variables

3.1 Soem

(and General

RandomConcepts

Vectors) 67

Here

1

x 1 e x , x > 0 > 0, > 0,

()

f x = ( )

0 , x0

where () = 0 y 1 e y dy (which exists and is finite for > 0). (This integral

is known as the Gamma function.) The distribution of X is also called the

Gamma distribution and , are called the parameters of the distribution.

Clearly, f(x) 0 and that f(x)dx = 1 is seen as follows.

f (x )dx = ( )

1 1

( )

x 1 e x dx = y 1 e y dy,

0

0

f (x )dx = ( ) 0 ( )

1 1

y 1 e y dy = = 1.

( )

REMARK 3 One easily sees, by integrating by parts, that

( ) (

= 1 1 , )( )

and if is an integer, then

( ) ( )(

= 1 2 1 , ) ()

where

() () ( )

1 = e y dy = 1; that is, a = 1 !

0

We often use this notation even if is not an integer, that is, we write

( ) ( )

= 1 ! = y 1 e y dy for > 0.

0

1 1

= ! = .

2 2

We have

1 (1 2 )

= y e y dy.

2 0

By setting

t2

y1 2 =

t

, so that y=

2

, dy = t dt , t 0, .( )

2

68 3 On Random Variables and Their Distributions

we get

1 1

= 2 e t t dt = 2 e t

2 2

2 2

dt = ;

2 0 t 0

that is,

1 1

= ! = .

2 2

From this we also get that

3 1 1

= = , etc.

2 2

2 2

Graphs of the p.d.f. of the Gamma distribution for selected values of and

are given in Figs. 3.6 and 3.7.

The Gamma distribution, and in particular its special case the Negative

Exponential distribution, discussed below, serve as satisfactory models for

f (x)

1.00

% ! 1, & ! 1

0.75

0.50

% ! 2, & ! 1

0.25 % ! 4, & ! 1

x

0 1 2 3 4 5 6 7 8

Figure 3.6 Graphs of the p.d.f. of the Gamma distribution for several values of , .

f (x)

1.00

0.50

% ! 2, & ! 1

0.25

% ! 2, & ! 2

x

0 1 2 3 4 5

Figure 3.7 Graphs of the p.d.f. of the Gamma distribution for several values of , .

3.3 Continuous Random Variables

3.1 Soem

(and General

RandomConcepts

Vectors) 69

well as in statistics.

For specific choices of the parameters and in the Gamma distribution,

we obtain the Chi-square and the Negative Exponential distributions given

below.

3.3.3 Chi-square

For = r/2, r 1, integer, = 2, we get what is known as the Chi-square

distribution, that is,

1 ( r 2 )1 x 2

x e , x>0

() 1

( )

f x = 2 r 2

r 2

r > 0, integer.

0, x0

The distribution with this p.d.f. is denoted by r2 and r is called the number of

degrees of freedom (d.f.) of the distribution. The Chi-square distribution occurs

often in statistics, as will be seen in subsequent chapters.

For = 1, = 1/, we get

e x , x>0

()

f x =

x0

> 0,

0 ,

which is known as the Negative Exponential distribution. The Negative Expo-

nential distribution occurs frequently in statistics and, in particular, in waiting-

time problems. More specifically, if X is an r.v. denoting the waiting time

between successive occurrences of events following a Poisson distribution,

then X has the Negative Exponential distribution. To see this, suppose that

events occur according to the Poisson distribution P(); for example, particles

emitted by a radioactive source with the average of particles per time unit.

Furthermore, we suppose that we have just observed such a particle, and let X

be the r.v. denoting the waiting time until the next particle occurs. We shall

show that X has the Negative Exponential distribution with parameter .

To this end, it is mentioned here that the distribution function F of an r.v., to

be studied in the next chapter, is defined by F(x) = P(X x), x !, and if X

is a continuous r.v., then dFdx( x ) = f ( x). Thus, it suffices to determine F here.

Since X 0, it follows that F(x) = 0, x 0. So let x > 0 be the the waiting time

for the emission of the next item. Then F(x) = P(X x) = 1 P(X > x). Since

is the average number of emitted particles per time unit, their average

x ( x )

0

= e x , since no

particles are emitted in (0, x]. That is, F(x) = 1 e , so that f(x) = ex. To

x

as asserted.

70 3 On Random Variables and Their Distributions

X NE(), or X r2 in order to denote the fact that X is distributed as

Gamma with parameters and , or Negative Exponental with parameter ,

or Chi-square with r degrees of freedom, respectively.

( ) (

X S = ! actually X S = , and ( ) [ ])

()

f x = (

1 , ) x

< .

0, otherwise

Clearly,

() f (x )dx = dx = 1.

1

f x 0,

are the parameters of the distribution. The interpretation of this distribution

is that subintervals of [, ], of the same length, are assigned the same prob-

ability of being observed regardless of their location. (See Fig. 3.8.)

The fact that the r.v. X has the Uniform distribution with parameters

and may be denoted by X U(, ).

f (x)

1

&"% Figure 3.8 Graph of the p.d.f. of the

U(, ) distribution.

x

0 % &

3.3.6 Beta

( ) (

X S = ! actually X S = 0, 1 and ( ) ( ))

+ ( )

( )

1

x 1 1 x 0<x<1

()

f x = ( )()

,

0 elsewhere, > 0, > 0.

Clearly, f(x) 0. That f ( x)dx = 1 is seen as follows.

3.3 Continuous Random Variables

3.1 Soem

(and General

RandomConcepts

Vectors) 71

( )()

= x 1 e x dx y 1 e y dy

0 0

( x+y )

= x 1 y 1 e dx dy

0 0

x=

uy

, dx =

y du

, u 0, 1 ( ) and x + y =

y

,

1 u

( ) 1 u

2

1 u

becomes

1 u 1 (

y 1 u ) du

= y 1 y 1 e y dy

(1 u) (1 u)

0 0 1 2

1 u 1 y 1 u( )

= y + 1 e du dy.

( )

0 0 +1

1 u

integral is

u (1 u)

1 1

= 1

+ 1 e du d

0 0

( )

1 1

= + 1 e d u 1 1 u du

0 0

( ) u (1 u)

1

1 1

= + du;

0

that is,

( )() ( ) x (1 x)

1

1 1

= + dx

0

and hence

(

+ )

f (x )dx = ( )( ) 0 x (1 x )

1 1

1

dx = 1.

Graphs of the p.d.f. of the Beta distribution for selected values of and are

given in Fig. 3.9.

REMARK 4 For = = 1, we get the U(0, 1), since (1) = 1 and (2) = 1.

The distribution of X is also called the Beta distribution and occurs rather

often in statistics. , are called the parameters of the distribution

and the function defined by 0 x 1 (1 x ) 1 dx for , > 0 is called the Beta

1

function.

Again the fact that X has the Beta distribution with parameters and

may be expressed by writing X B(, ).

72 3 On Random Variables and Their Distributions

f (x) %!5

&!3

2.5

%!3

&!3

2.0

%!2

1.5

&!2

1.0

x

0 0.2 0.4 0.6 0.8 1.0

Figure 3.9 Graphs of the p.d.f. of the Beta distribution for several values of , .

3.3.7 Cauchy

Here

( )

X S = ! and ()

f x =

1

, x ! , ! , > 0.

2 + x

( )

2

f ( x )dx =

1 1 1

1 + x

dx = dx

( ) [( ) ]

2 2

2 + x

1 dy 1

1 + y 2

= = arctan y = 1,

upon letting

x dx

y= , so that = dy.

The distribution of X is also called the Cauchy distribution and , are called

the parameters of the distribution (see Fig. 3.10). We may write X

Cauchy(, 2) to express the fact that X has the Cauchy distribution with

parameters and 2.

(The p.d.f. of the Cauchy distribution looks much the same as the Normal

p.d.f., except that the tails of the former are heavier.)

3.3.8 Lognormal

Here X(S) = ! (actually X(S) = (0, )) and

3.3 3.1 Soem

Continuous Random Variables (and General

RandomConcepts

Vectors) 73

f (x)

0.3

0.2

0.1

x

"2 "1 0 1 2

Figure 3.10 Graph of the p.d.f. of the Cauchy distribution with = 0, = 1.

( )

2

1 exp log x log , x>0

()

f x = x 2

2 2

0, x 0 where , > 0.

Now f(x) 0 and

log x log

( )

2

()

1 1

f x dx = 0 exp dx

2 x 2 2

which, letting x = ey, so that log x = y, dx = eydy, y (, ), becomes

y log

( )

2

1 1

= y 2 2

2 e

exp e y dy.

But this is the integral of an N(log , 2) density and hence is equal to 1; that

is, if X is lognormally distributed, then Y = log X is normally distributed with

parameters log and 2. The distribution of X is called Lognormal and , are

called the parameters of the distribution (see Fig. 3.11). The notation X

Lognormal(, ) may be used to express the fact that X has the Lognormal

distribution with parameters and .

(For the many applications of the Lognormal distribution, the reader is

referred to the book The Lognormal Distribution by J. Aitchison and J. A. C.

Brown, Cambridge University Press, New York, 1957.)

"

These distributions occur very often in Statistics (interval esti-

3.3.9 t

mation, testing hypotheses, analysis of variance, etc.) and their

3.3.10 F densities will be presented later (see Chapter 9, Section 9.2).

We close this section with an example of a continuous random vector.

Here X(S) = !2 (that is, X is a 2-dimensional random vector) with

74 3 On Random Variables and Their Distributions

f (x)

0.8 & 2 ! 0.5

%!1

0.6

1

0.4 % ! e2

0.2 %!e

x

0 1 2 3 4

Figure 3.11 Graphs of the p.d.f. of the Lognormal distribution for several values of , .

(

f x1 , x 2 = ) 1

e q 2 ,

2 1 2 1 2

1 x 2 x 1 1 x 2 2 x 2 2

2

q= 1 1

2 +

1 2 1 1 2 2

with 1, 2 !. The distribution of X is also called the Bivariate Normal

distribution and the quantities 1, 2, 1, 2, are called the parameters of the

distribution. (See Fig. 3.12.)

Clearly, f(x1, x2) > 0. That ! f(x1, x2)dx1dx2 = 1 is seen as follows:

2

x 2 x1 1 x 2 2 x 2 2

2

( )

1 2 q = 1

1

1

2

1 2 2

+

2 2

x 2 x1 1

= 2

2

1 1

(

2 x1 1

+ 1 . )

Furthermore,

x2 2 x1 1 x 2 2 1 x 1

= 2 1

2 1 2 2 1

x1 1

=

1

2

x 2 2 + 2

=

1

x2 b , ( )

1 2

where

2

b = 2 +

1

(

x1 1 . )

3.3 3.1 Soem

Continuous Random Variables (and General

RandomConcepts

Vectors) 75

z

z ! f(x1, x 2 ) for z ' k

x2

0

x1 k

Thus

2 2

(

1 2 q = 2)

x b

2

2 x1 1

+ 1

1

( )

and hence

x

(

)

2

( )

1

exp

1 1

f x1 , x 2 dx 2 =

2 1 2 12

( )

2

1 x2 b dx

exp

2 2 1 2

2 2 1 2

2 ( ) 2

x

( )

2

1

exp 1,

1 1

=

2 1 2 2

1

since the integral above is that of an N(b, 22(1 2)) density. Since the first

factor is the density of an N(1, 12) random variable, integrating with respect

to x1, we get

f ( x1 , x 2 )dx1 dx 2 = 1.

REMARK 5 From the above derivations, it follows that, if f(x1, x2) is Bivariate

Normal, then

( ) ( ) ( )

f1 x1 = f x1 , x 2 dx 2 is N 1 , 12 ,

and similarly,

( ) ( ) ( )

f2 x 2 = f x1 , x 2 dx1 is N 2 , 22 .

76 3 On Random Variables and Their Distributions

As will be seen in Chapter 4, the p.d.f.s f1 and f2 above are called marginal

p.d.f.s of f.

The notation X N(1, 2, 12, 22, ) may be used to express the fact that

X has the Bivariate Normal distribution with parameters 1, 2, 12, 22, .

Then X1 N(1 , 12) and X2 N(2, 22 ).

Exercises

3.3.1 Let f be the p.d.f. of the N(, 2) distribution and show that:

ii) f is symmetric about ;

ii) max f x = () 1

.

x !

2

3.3.2 Let X be distributed as N(0, 1), and for a < b, let p = P(a < X < b). Then

use the symmetry of the p.d.f. f in order to show that:

iii) For 0 a < b, p = (b) (a);

iii) For a 0 < b, p = (b) + (a) 1;

iii) For a b < 0, p = (a) (b);

iv) For c > 0, P(c < X < c) = 2(c) 1.

(See Normal Tables in Appendix III for the definition of .)

3.3.3 If X N(0, 1), use the Normal Tables in Appendix III in order to show

that:

iii) P(1 < X < 1) = 0.68269;

iii) P(2 < X < 2) = 0.9545;

iii) P(3 < X < 3) = 0.9973.

3.3.4 Let X be a r2 . In Table 5, Appendix III, the values = P(X x) are

given for r ranging from 1 to 45, and for selected values of . From the entries

of the table, observe that, for a fixed , the values of x increase along with the

number of degrees of freedom r. Select some values of and record the

corresponding values of x for a set of increasing values of r.

3.3.5 Let X be an r.v. distributed as 102. Use Table 5 in Appendix III in order

to determine the numbers a and b for which the following are true:

ii) P(X < a) = P(X > b);

ii) P(a < X < b) = 0.90.

3.3.6 Consider certain events which in every time interval [t1, t2] (0 < t1 < t2)

occur independently for nonoverlapping intervals according to the Poisson

distribution P((t2 t1)). Let T be the r.v. denoting the time which lapses

3.1 Soem General Exercises

Concepts 77

between two consecutive such events. Show that the distribution of T is Nega-

tive Exponential with parameter by computing the probability that T > t.

3.3.7 Let X be an r.v. denoting the life length of a TV tube and suppose that

its p.d.f. f is given by:

f(x) = exI(0, )(x).

Compute the following probabilities:

iii)P(j < X j + 1), j = 0, 1, . . . ;

iii)P(X > t) for some t > 0;

iii)P(X > s + t|X > s) for some s, t > 0;

iv) Compare the probabilities in parts (ii) and (iii) and conclude that the

Negative Exponential distribution is memoryless;

iv) If it is known that P(X > s) = , express the parameter in terms of

and s.

3.3.8 Suppose that the life expectancy X of each member of a certain group

of people is an r.v. having the Negative Exponential distribution with param-

eter = 1/50 (years). For an individual from the group in question, compute

the probability that:

iii) He will survive to retire at 65;

iii) He will live to be at least 70 years old, given that he just celebrated his 40th

birthday;

iii) For what value of c, P(X > c) = 12 ?

3.3.9 Let X be an r.v. distributed as U(, ) ( > 0). Determine the values

of the parameter for which the following are true:

ii) P(1 < X < 2) = 0.75;

ii) P(|X| < 1) = P(|X| > 2).

3.3.10 Refer to the Beta distribution and set:

1

( ) ( )

1

B , = x 1 1 x dx.

0

3.3.11 Establish the following identity:

n 1 p m1

( ) n!

( )

n m p n m

n x 1 x dx = x m1 1 x dx

m 1 0 ( )(

m1 nm ! 0 )

n

= n p j 1 p ( )

n j

.

j= m

j

3.3.12 Let X be an r.v. with p.d.f given by f(x) = 1/[(1 + x2)]. Calculate the

probability that X2 c.

78 3 On Random Variables and Their Distributions

= xex /2I(0, )(x) (Raleigh distribution);

2

iii) f(x)

= 2 / x2ex /2I(0,)(x) (Maxwells distribution);

2

iii) f(x)

iii) f(x) = 12 e|x| (Double Exponential);

iv) f(x) = ( ca )( cx )+1IA(x), A = (c, ), , c > 0 (Pareto distribution).

ii) f(x) = cos xI(0,/2)(x);

ii) f(x) = xexI(0,)(x).

3.3.15 For what values of the constant c are the following functions p.d.f.s?

ce 6 x , x>0

ii) f(x) = cx, 1 < x 0;

0 , x 1

ii) f(x) = cx2ex I(0,) (x).

3

3.3.16 Let X be an r.v. with p.d.f. given by 3.3.15(ii). Compute the probabil-

ity that X > x.

3.3.17 Let X be the r.v. denoting the life length of a certain electronic device

expressed in hours, and suppose that its p.d.f. f is given by:

()

f x =

c

()

I 1,000 , 3,000 ] x .

xn [

ii) Calculate the probability that the life span of one electronic device of the

type just described is at least 2,000 hours.

3.3.18 Refer to Exercise 3.3.15(ii) and compute the probability that X ex-

ceeds s + t, given that X > s. Compare the answer with that of Exercise

3.3.7(iii).

3.3.19 Consider the function f(x) = x 1e x , x > 0 (, > 0), and:

iii) Show that it is a p.d.f. (called the Weibull p.d.f. with parameters

and );

iii) Observe that the Negative Exponential p.d.f. is a special case of a Weibull

p.d.f., and specify the values of the parameters for which this happens;

iii) For = 1 and = 12 , = 1 and = 2, draw the respective graphs of the

p.d.f.s involved.

(Note: The Weibull distribution is employed for describing the lifetime of

living organisms or of mechanical systems.)

3.4

3.1 The

SoemPoisson

General

Distribution

Concepts 79

3.3.20 Let X and Y be r.v.s having the joint p.d.f. f given by:

( ) ( ) (

f x, y = c 25 x 2 y 2 I (0 ,5 ) x 2 + y 2 . )

Determine the constant c and compute the probability that 0 < X2 + Y2 < 4.

3.3.21 Let X and Y be r.v.s whose joint p.d.f. f is given by f(x, y) =

cxyI(0,2)(0,5)(x, y). Determine the constant c and compute the following

probabilities:

i) P( 12 < X < 1, 0 < Y < 3);

ii) P(X < 2, 2 < Y < 4);

iii) P(1 < X < 2, Y > 5);

iv) P(X > Y).

3.3.22 Verify that the following function is a p.d.f.:

( )

f x, y =

1

4

( ) (

cos y I A x, y , ) ( ]

A = , , .

2 2

3.3.23 (A mixed distribution) Show that the following function is a p.d.f.

1 x

e , x0

4

1 , 0<x<2

()

f x = 8

x

1

, x = 2, 3,

2

0, otherwise.

the Binomial Distribution and the Binomial Distribution as

an Approximation to the Hypergeometric Distribution

In this section, we first establish rigorously the assertion made earlier that the

Poisson distribution may be obtained as the limit of Binomial distributions. To

this end, consider a sequence of Binomial distributions, so that the nth distri-

bution has probability pn of a success, and we assume that as n , pn 0 and

that

n = npn ,

for some > 0. Then the following theorem is true.

THEOREM 1 With the above notation and assumptions, we have

x

n x n x

pn q n n

e for each fixed x = 0, 1, 2, .

x

x!

80 3 On Random Variables and Their Distributions

PROOF We have

( ) (

n x n x n n 1 n x + 1 x n x

pn q n = pn q n

)

x x!

=

( ) ( )

n n 1 n x + 1 n n

x n x

1

x! n n

=

( ) (

n n1 nx +1 ) 1 x

n

1 n

1

n

nx x! n n

x

1

n

1 x 1 nx

= 1 1 1

n n x!

n

1 x

1 n

n

e ,

n n

x

x!

1

n

since, if n , then

n

n

1 e .

n

n

1 e .

n

probabilities ( nx )pxqnx are not easily calculated. Then we can approximate

them by e(x/x!), provided p is small, where we replace be np. This is true

for all x = 0, 1, 2, . . . , n.

We also meet with difficulty in calculating probabilities in the Hyper-

geometric distribution

m n m + n

x r x r

if m, n are large. What we do is approximate the Hypergeometric distribution

by an appropriate Binomial distribution, and then, if need be, we can go one

step further in approximating the Binomial by the appropriate Poisson distri-

bution according to Theorem 1. Thus we have

THEOREM 2 Let m, n and suppose that m/(m + n) = pm,n p, 0 < p < 1. Then

3.4

3.1 The

SoemPoisson

General

Distribution

Concepts 81

m n

x r x r

p x q r x , x = 0, 1, 2, , r .

m + n x

r

PROOF We have

m n

x r x

=

(

m! n! m + n r ! )

r!

m + n

( )[ ( )] ( )

m x ! n r x ! m + n ! x! r x ! ( )

r

( ) (

r m x + 1 m n r + x + 1 n

=

)

x (

m+nr +1 m+n) ( )

r m( m 1) [ m ( x 1)] n( n 1) [ n (r x 1)]

= .

x (m + n) [(m + n) (r 1)]

Both numerator and denominator have r factors. Dividing through by (m + n),

we get

m x

n r x r m m 1 m x1

=

m + n x m + n m + n m + n m + n m + n

r

n n 1 n r x 1

m + n m + n m + n m+n m+n

1

1 r 1

1 1 1

m + n m + n

r

m ,n

p x q r x ,

x

since

m n

p and hence 1 p = q. "

m+n m+n

approximate the probabilities

82 3 On Random Variables and Their Distributions

m n

x r x r

by p x q r x

m + n x

r

by setting p = m/(m +x n). This is true for all x = 0, 1, 2, . . . , r. It is to be

observed that ( xr )( mm+ n ) (1 mm+ n ) is the exact probability of having exactly x

rx

probability of a success, mm+ n , remains constant. The Hypergeometric distribu-

tion is appropriate when sampling is done without replacement. If, however,

m

m and n are large (m, n ) and n remains approximately constant

( mn c = p / q), then the probabilities of having exactly x successes in r

(independent) trials are approximately equal under both sampling schemes.

Exercises

3.4.1 For the following values of n, p and = np, draw graphs of B(n, p) and

P() on the same coordinate axes:

iii) n= 10, p = 4

16

, so that = 2.5;

iii) n= 16, p = 2

16

, so that = 2;

iii) n= 20, p = 2

16

, so that = 2.5;

iv) n= 24, p = 1

16

, so that = 1.5;

iv) n= 24, p = 2

16

, so that = 3.

3.4.2 Refer to Exercise 3.2.2 and suppose that the number of applicants is

equal to 72. Compute the probabilities (i)(iii) by using the Poisson approxi-

mation to Binomial (Theorem 1).

3.4.3 Refer to Exercise 3.2.10 and use Theorem 2 in order to obtain an

approximate value of the required probability.

In this section, random variables and random vectors are introduced as special

cases of measurable functions. Certain results related to -fields are also

derived. Consider the probability space (S, A, P) and let T be a space and X

be a function defined on S into T, that is, X : S T. For T T, define the

inverse image of T, under X, denoted by X1(T), as follows:

( ) { () }

X 1 T = s S ; X s T .

3.5* Random Variables as Measurable Functions

3.1 Soem and

General

Related

Concepts

Results 83

properties are immediate consequences of the definition (and the fact X is a

function):

j j

( )

X 1 U T j = U X 1 T j . (1)

( ) ( )

If T1 T2 = , then X 1 T1 X 1 T2 = . (2)

Hence by (1) and (2) we have

j j

( )

X 1 T j = X 1 T j . (3)

Also

j j

( )

X 1 I T j = I X 1 T j , (4)

( ) [ ( )] ,

c

X 1 T c = X 1 T (5)

( )

X 1 T = S , (6)

X 1 () = . (7)

subsets of S as follows:

( ) { ( )

X 1 D = A S ; A = X 1 T for some T D . }

By means of (1), (5), (6) above, we immediately have

THEOREM 3 The class X1(D) is a -field of subsets of S.

The above theorem is the reason we require measurability in our defini-

tion of a random variable. It guarantees that the probability distribution

function of a random vector, to be defined below, is well defined.

If X1(D) A, then we say that X is (A, D)-measurable, or just measur-

able if there is no confusion possible. If (T, D) = (!, B) and X is (A, B)-

measurable, we say that X is a random variable (r.v.). More generally, if (T, D)

= (!k, Bk), where !k = ! ! ! (k copies of !), and X is (A, Bk)-

measurable, we say that X is a k-dimensional random vector (r. vector). In this

latter case, we shall write X if k 1, and just X if k = 1. A random variable is

a one-dimensional random vector.

On the basis of the properties (1)(7) of X1, the following is immediate.

THEOREM 4 Define the class C* of subsets of T as follows: C* = {T T; X1(T) = A for some

A A}. Then C* is a -field.

COROLLARY Let D = (C), where C is a class of subsets of T. Then X is (A, D)-measurable

if and only if X1(C) A. In particular, X is a random variable if and only if

84 3 On Random Variables and Their Distributions

of k-dimensional random vectors. The classes Co, Cj, C j, j = 1, . . . , 8 are

defined in Theorem 5 and the paragraph before it in Chapter 1.

PROOF The -field C* of Theorem 4 has the property that C* C. Then C*

D = (C) and hence X1(C*) X1(D). But X1(C*) A. Thus X1(D)

A. The converse is a direct consequence of the definition of (A, D) =

measurability.

Now, by means of an r. vector X: (S, A, P) (!k, Bk ), define on Bk the

set function PX as follows:

( ) [ ( )] ( ) ({

PX B = P X 1 B = P X B = P s S ; X s B . ( ) }) (8)

By the Corollary to Theorem 4, the sets X1(B) in S are actually events due to

the assumption that X is an r. vector. Therefore PX is well defined by (8); i.e.,

P[X1(B)] makes sense, is well defined. It is now shown that PX is a probability

measure on Bk . In fact, PX(B) 0, B Bk, since P is a probability measure.

Next, PX(!k) = P[X 1(!k)] = P(S ) = 1, and finally,

[ ( )]

( ) ( )

PX B j = P X 1 B j = P X 1 B j = P X 1 B j = PX B j .

j =1 j = 1

j =1 j =1 j =1

just the distribution) of X.

Exercises

3.5.1 Consider the sample space S supplied with the -field of events A. For

an event A, the indicator IA of A is defined by: IA(s) = 1 if s A and IA(s) = 0

if s Ac.

iii) Show that IA is r.v. for any A A .

iii) What is the partition of S induced by IA?

iii) What is the -field induced by IA?

3.5.2 Write out the proof of Theorem 1 by using (1), (5) and (6).

3.5.3 Write out the proof of Theorem 2.

4.1 The Cumulative Distribution Function 85

Chapter 4

Densities, and Their Relationship

Basic Properties of the d.f. of a Random Variable

The distribution of a k-dimensional r. vector X has been defined through the

relationship: PX (B) = P(X B), where B is a subset of ! k. In particular, one

may choose B to be an interval in ! k; i.e., B = {y ! k; y x} in the sense that,

if x = (x1, . . . , xk) and y = (y1, . . . , yk), then yj xj, j = 1, . . . , k. For such a choice

of B, PX(B) is denoted by FX(x) and is called the cumulative distribution

function of X (evaluated at x), or just the distribution function (d.f.) of X. We

omit the subscript X if no confusion is possible. Thus, the d.f. F of a r. vector

X is an ordinary point function defined on ! k (and taking values in [0, 1]). Now

we restrict our attention to the case k = 1 and prove the following basic

properties of the d.f. of an r.v.

THEOREM 1 The distribution function F of a random variable X satisfies the following

properties:

i) 0 F(x) 1, x !.

ii) F is nondecreasing.

iii) F is continuous from the right.

iv) F(x) 0 as x , F(x) 1, as x +.

We express this by writing F() = 0, F(+) = 1.

PROOF In the course of this proof, we set Q for the distribution PX of X, for

the sake of simplicity. We have then:

i) Obvious.

ii) This means that x1 < x2 implies F(x1) F(x2). In fact,

85

86 4 Distribution Functions, Probability Densities, and Their Relationship

x1 < x2 implies (, x ] (, x ]

1 2

and hence

( ] ( ]

Q , x1 Q , x 2 ; equivalently, F x1 F x 2 . ( ) ( )

iii) This means that, if xn x, then F(xn) F(x). In fact,

x n x implies (, x ] (, x]

n

and hence

( ]

Q , x n Q , x ( ]

by Theorem 2, Chapter 2; equivalently, F(xn) F(x).

iv) Let xn . We may assume that xn (see also Exercise 4.1.6). Then

(, x ] ,

n (

so that Q , x n Q = 0] ( )

by Theorem 2, Chapter 2. Equivalently, F(xn) 0. Similarly, if xn +.

We may assume xn . Then

(, x ] ! and hence Q(, x ] Q(! ) = 1; equivalently, F ( x ) 1. "

n n n

REMARK 1

i) F(x) can be used to find probabilities of the form P(a < X b); that is

(

P a<X b =F b F a . ) () ()

In fact,

(a < X b) = ( < X b) ( < X a)

and

( < X a) ( < X b).

Thus

( ) ( ) (

P a < X b = P < X b P < X a = F b F a . ) () ()

ii) The limit from the left of F(x) at x, denoted by F(x), is defined as follows:

( )

F x = lim F x n

n

( ) with x n x.

This limit always exists, since F(xn), but need not be equal to F(x+)(=limit

from the right) = F(x). The quantities F(x) and F(x) are used to express

the probability P(X = a); that is, P(X = a) = F(a) F(a). In fact, let xn

a and set A = (X = a), An = (xn < X a). Then, clearly, An A and hence

by Theorem 2, Chapter 2,

( ) ( )

P An P A , or lim P x n < X a = P X = a ,

n

( ) ( )

or

[ ( ) ( )] (

lim F a F x n = P X = a ,

n

)

4.1 The Cumulative Distribution Function 87

or

() ( )

F a lim F x n = P X = a ,

n

( )

or

() ( )

F a F a =P X =a . ( )

It is known that a nondecreasing function (such as F) may have

discontinuities which can only be jumps. Then F(a) F(a) is the length of

the jump of F at a. Of course, if F is continuous then F(x) = F(x) and

hence P(X = x) = 0 for all x.

iii) If X is discrete, its d.f. is a step function, the value of it at x being defined

by

( ) f (x )

F x =

xjx

j and ( ) ( ) ( )

f x j = F x j F x j1 ,

iv) If X is of the continuous type, its d.f. F is continuous. Furthermore,

dF x ( )=f

dx

(x)

F(x) F(x)

1.00 1.00

0.80 0.80

0.60 0.60

0.40 0.40

0.20 0.20

0 x 0 x

1

(a) Binomial for n ! 6, p ! 4

. (b) Poisson for ( ! 2.

F(x) )(x)

1.0 1.0

0 x 0.5

% &

x

0 x+% "2 "1 0 1 2

x"%

(c) U (%, &). Here F(x) ! &"%

% * x * &.

(d) N(0, 1).

1 x'&

Figure 4.1 Examples of graphs of c.d.f.s.

88 4 Distribution Functions, Probability Densities, and Their Relationship

relations

dF x ( )=f

() () (x),

x

F x = f t dt and

dx

we see that if f is continuous, f determines F( f F) and F determines f (F

f ); that is, F f. Two important applications of this are the following two

theorems.

Often one has to deal with functions of an r. vector itself. In such cases, the

resulting entities have got to be r. vectors, since we operate in a probability

framework. The following statement is to this effect. Its precise formulation

and justification is given as Theorem 7 on page 104.

STATEMENT 1 Let X be a k-dimensional r. vector defined on the sample space S, and let g be

a (well-behaving) function defined on ! k and taking values in ! m. Then g(X)

is defined on the underlying sample space S, takes values in ! m, and is an r.

vector. (That is, well-behaving functions of r. vectors are r. vectors.) In particu-

lar, g(X) is an r. vector if g is continuous.

Now a k-dimensional r. vector X may be represented in terms of its

coordinates; i.e., X = (X1, . . . , Xk), where Xj, j = 1, . . . , k are real-valued

functions defined on S. The question then arises as to how X and Xj, j = 1, . . . ,

k are related from the point of view of being r. vectors. The answer is provided

by the following statement, whereas the precise statement and justification are

given as Theorem 8 below.

STATEMENT 2 Let X and Xj, j = 1, . . . , k be functions defined on the sample space S and

taking values in ! k and !, respectively, and let X = (X1, . . . , Xk). Then X is an

r. vector if and only if Xj, j = 1, . . . , k are r.v.s.

The following two theorems provide applications of Statement 1.

THEOREM 2 Let X be an N(, 2)-distributed r.v. and set Y = X

. Then Y is an r.v. and its

distribution is N(0, 1).

PROOF In the first place, Y is an r.v. by Statement 1. Then it suffices to show

that the d.f. of Y is , where

() 1 y

2

y = e t 2

dt .

2

We have

X

( )

P Y y = P

y = P X y +

( )

(

t

)

2

1 1

()

y +

exp dt = y

e

u2 2

= du = y ,

2 2 2 2

4.1 The Cumulative DistributionExercises

Function 89

REMARK 2 The transformation X

of X is referred to as normalization

of X.

THEOREM 3 i(i) Let X be an N(0, 1)-distributed r.v. Then Y = X 2 is distributed as 21.

(ii) If X is a N ( , 2)-distributed r.v., then the r.v. ( X )2 is distributed as 21.

PROOF (i) We will show that the p.d.f. of Y is that of a 21-distributed r.v. by

deriving the d.f. of Y first and then differentating it in order to obtain fY. To this

end, let us observe first that Y is an r.v. on account of Statement 1. Next, for

y > 0, we have

() ( )

FY y = P Y y = P y X ( y )

1 y 1 y

2 2

= ex 2

dx = 2 ex 2

dx.

y

2 2 0

() 1 1

y

FY y = 2

2

0

2 t

e t 2 dt .

Hence

dFY y( )= 1 1

e y 2

=

1

y

(1 2 )1

e y 2 .

dy 2 y 2

1 (1 2 )1 y 2

y e , y>0

()

fY y = 21 2

0, y 0,

and this is the p.d.f. of 21. (Observe that here we used the fact that ( 12 ) =

.) "

Exercises

4.1.1 Refer to Exercise 3.2.13, in Chapter 3, and determine the d.f.s corre-

sponding to the p.d.f.s given there.

4.1.2 Refer to Exercise 3.2.14, in Chapter 3, and determine the d.f.s corre-

sponding to the p.d.f.s given there.

90 4 Distribution Functions, Probability Densities, and Their Relationship

4.1.3 Refer to Exercise 3.3.13, in Chapter 3, and determine the d.f.s corre-

sponding to the p.d.f.s given there.

4.1.4 Refer to Exercise 3.3.14, in Chapter 3, and determine the d.f.s corre-

sponding to the p.d.f.s given there.

4.1.5 Let X be an r.v. with d.f. F. Determine the d.f. of the following r.v.s:

X, X 2, aX + b, XI[a,b) (X ) when:

i) X is continuous and F is strictly increasing;

ii) X is discrete.

4.1.6 Refer to the proof of Theorem 1 (iv) and show that we may assume

that xn (xn ) instead of xn (xn ).

4.1.7 Let f and F be the p.d.f. and the d.f., respectively, of an r.v. X. Then

show that F is continuous, and dF(x)/dx = f(x) at the continuity points x of f.

4.1.8

i) Show that the following function F is a d.f. (Logistic distribution) and

derive the corresponding p.d.f., f.

()

F x =

1

(x + )

, x ! , > 0, ! ;

1+ e

ii) Show that f(x) =F(x)[1 F(x)].

4.1.9 Refer to Exercise 3.3.17 in Chapter 3 and determine the d.f. F corre-

sponding to the p.d.f. f given there. Write out the expressions of F and f for

n = 2 and n = 3.

4.1.10 If X is an r.v. distributed as N(3, 0.25), use Table 3 in Appendix III in

order to compute the following probabilities:

i) P(X < 1);

ii) P(X > 2.5);

iii) P(0.5 < X < 1.3).

4.1.11 The distribution of IQs of the people in a given group is well approxi-

mated by the Normal distribution with = 105 and = 20. What proportion of

the individuals in the group in question has an IQ:

i) At least 150?

ii) At most 80?

iii) Between 95 and 125?

4.1.12 A certain manufacturing process produces light bulbs whose life

length (in hours) is an r.v. X distributed as N(2,000, 2002). A light bulb is

supposed to be defective if its lifetime is less than 1,800. If 25 light bulbs are

4.2 The 4.1

d.f. ofThe

a Random

Cumulative

Vector

Distribution

and Its Properties

Function 91

tested, what is the probability that at most 15 of them are defective? (Use the

required independence.)

4.1.13 A manufacturing process produces 12 -inch ball bearings, which are

assumed to be satisfactory if their diameter lies in the interval 0.5 0.0006 and

defective otherwise. A days production is examined, and it is found that the

distribution of the actual diameters of the ball bearings is approximately

normal with mean = 0.5007 inch and = 0.0005 inch. Compute the propor-

tion of defective ball bearings.

4.1.14 If X is an r.v. distributed as N(, 2), find the value of c (in terms of

and ) for which P(X < c) = 2 9P(X > c).

4.1.15 Refer to the Weibull p.d.f., f, given in Exercise 3.3.19 in Chapter 3 and

do the following:

i) Calculate the corresponding d.f. F and the reliability function !(x) = 1

F(x);

()

f (x )

ii) Also, calculate the failure (or hazard) rate H x = ! ( x ) , and draw its graph

for = 1 and = 1 , 1, 2;

2

iii) For s and t > 0, calculate the probability P(X > s + t|X > t) where X is an r.v.

having the Weibull distribution;

iv) What do the quantities F(x), !(x), H(x) and the probability in

part (iii) become in the special case of the Negative Exponential

distribution?

4.2 The d.f. of a Random Vector and Its PropertiesMarginal and Conditional

d.f.s and p.d.f.s

For the case of a two-dimensional r. vector, a result analogous to Theorem

1 can be established. So consider the case that k = 2. We then have X =

(X1, X2) and the d.f. F(or FX or FX1,X2) of X, or the joint distribution function

of X1, X2, is F(x1, x2) = P(X1 x1, X2 x2). Then the following theorem holds

true.

With the above notation we have

THEOREM 4

i) 0 F(x1, x2) 1, x1, x2 !.

ii) The variation of F over rectangles with sides parallel to the axes, given in

Fig. 4.2, is 0.

iii) F is continuous from the right with respect to each of the coordinates x1, x2,

or both of them jointly.

92 4 Distribution Functions, Probability Densities, and Their Relationship

y

(x1, y 2)

y2 (x2, y 2)

" ,

Figure 4.2 The variation V of F over the

rectangle is:

, "

y1 (x2, y 1) F(x1, y1) + F(x2, y2) F(x1, y2) F(x2, y1)

(x1, y 1)

x

0 x1 x2

iv) If both x1, x2, , then F(x1, x2) 1, and if at least one of the x1, x2

, then F(x1, x2) 0. We express this by writing F(, ) = 1, F(, x2) =

F(x1, ) = F(, ) = 0, where < x1, x2 < .

PROOF

i) Obvious.

ii) V = P(x1 < X1 x2, y1 < X2 y2) and is hence, clearly, 0.

iii) Same as in Theorem 3. (If x = (x1, x2), and zn = (x1n, x2n), then zn x means

x1n x1, x2n x2).

iv) If x1, x2 , then (, x1] (, x2] R 2, so that F(x1, x2) P(S) = 1. If

at least one of x1, x2 goes () to , then (, x1] (, x2] , hence

( ) ( )

F x1 , x 2 P = 0. "

REMARK 3 The function F(x1, ) = F1(x1) is the d.f. of the random variable

X1. In fact, F(x1, ) = F1(x1) is the d.f. of the random variable X1. In fact,

( ) (

F x1 , = lim P X1 x1 , X 2 x n

x n

)

( ) (

= P X1 x1 , < X 2 < = P X1 x1 = F1 x1 . ) ( )

Similarly F(, x2) = F2(x2) is the d.f. of the random variable X2. F1, F2 are called

marginal d.f.s.

REMARK 4 It should be pointed out here that results like those discussed

in parts (i)(iv) in Remark 1 still hold true here (appropriately interpreted).

In particular, part (iv) says that F(x1, x2) has second order partial derivatives

and

2

x1x 2

( ) (

F x1 , x 2 = f x1 , x 2 )

at continuity points of f.

For k > 2, we have a theorem strictly analogous to Theorems 3 and 6 and

also remarks such as Remark 1(i)(iv) following Theorem 3. In particular, the

analog of (iv) says that F(x1, . . . , xk) has kth order partial derivatives and

4.2 The 4.1

d.f. ofThe

a Random

Cumulative

Vector

Distribution

and Its Properties

Function 93

k

x1x2 xk

(

F x1 , , xk = f x1 , , xk ) ( )

at continuity points of f, where F, or FX, or FX1, , Xk, is the d.f. of X, or the joint

distribution function of X1, . . . , Xk. As in the two-dimensional case,

(

F , , , x j , , , = F j x j ) ( )

is the d.f. of the random variable Xj, and if m xjs are replaced by (1 < m < k),

then the resulting function is the joint d.f. of the random variables correspond-

ing to the remaining (k m) Xjs. All these d.f.s are called marginal distribu-

tion functions.

In Statement 2, we have seen that if X = (X1, . . . , Xk) is an r. vector, then

Xj, j = 1, 2, . . . , k are r.v.s and vice versa. Then the p.d.f. of X, f(x) =

f(x1, . . . , xk), is also called the joint p.d.f. of the r.v.s X1, . . . , Xk.

Consider first the case k = 2; that is, X = (X1, X2), f(x) = f(x1, x2) and set

f x1 , x 2

( )

( )

f1 x1 = x 2

( )

f x1 , x 2 dx 2

f x1 , x 2

( )

( )

f2 x 2 = x 1

( )

f x1 , x 2 dx1 .

f ( x ) = f ( x , x ) = 1,

1 1 1 2

x1 x1 x2

or

Similarly we get the result for f2. Furthermore, f1 is the p.d.f. of X1, and f2 is the

p.d.f. of X2. In fact,

(

f x1 , x2 = f x1 , x2 = f1 x1

) ( ) ( )

( )

P X 1 B = x B , x !

1 2 x B x ! x B 1 2 1

( )

f x1 , x2 dx1dx2 = f x1 , x2 dx2 dx1 = f1 x1 dx1 .

B ! B ! B [ ( ) ] ( )

Similarly f2 is the p.d.f. of the r.v. X2. We call f1, f2 the marginal p.d.f.s. Now

suppose f1(x1) > 0. Then define f(x2|x1) as follows:

) (f (x ) ) .

f x,x

(

f x 2 x1 =

1

1

1

2

94 4 Distribution Functions, Probability Densities, and Their Relationship

X1 ( f1(x1) > 0). Then f(|x1) is a p.d.f. In fact, f(x2|x1) 0 and

f ( x 2 x1 ) = f ( x1 , x 2 ) = ( )

1 1

f1 x1 = 1,

x2 ( )

f1 x1 x2 ( )

f1 x1

f ( x 2 x1 )dx 2 = f ( x ) f ( x1 , x 2 )dx 2 = f ( x ) f1 ( x1 ) = 1.

1 1

1 1 1 1

) (f (x ) )

f x, x

(

f x1 x 2 =

2

1

2

2

and show that f(|x2) is a p.d.f. Furthermore, if X1, X2 are both discrete, the

f(x2|x1) has the following interpretation:

) (f (x ) ) = ( P(X = x ) ) = P(X

f x,x P X =x, X =x

(

f x 2 x1 =

1

1

1

2 1 1

1

2

1

2

2 = x 2 X1 = x1 . )

Hence P(X2 B|X1 = x1) = x B f(x2|x1). For this reason, we call f(|x2) the

2

conditional p.d.f. of X2, given that X1 = x1 (provided f1(x1) > 0). For a similar

reason, we call f(|x2) the conditional p.d.f. of X1, given that X2 = x2 (provided

f2(x2) > 0). For the case that the p.d.f.s f and f2 are of the continuous type,

the conditional p.d.f. f (x1|x2) may be given an interpretation similar to the

one given above. By assuming (without loss of generality) that h1, h2 > 0, one

has

1 1 1 1 1 2 2 2 2

=

(1 h h )P(x < X x + h , x < X x + h )

1 2 1 1 1 1 2 2 2 2

(1 h )P(x < X x + h )

2 2 2 2 2

=

(1 h h )[F (x , x ) + F (x + h , x + h ) F (x , x + h ) F (x

1 2 1 2 1 1 2 2 1 2 2 1 + h1 , x 2 )]

(1 h )[F (x + h ) F (x )] 2 2 2 2 2 2

where F is the joint d.f. of X1, X2 and F2 is the d.f. of X2. By letting h1, h2 0

and assuming that (x1, x2) and x2 are continuity points of f and f2, respectively,

the last expression on the right-hand side above tends to f(x1, x2)/f2(x2) which

was denoted by f(x1|x2). Thus for small h1, h2, h1 f(x1|x2) is approximately equal

to P(x1 < X1 x1 + h1|x2 < X2 x2 + h2), so that h1 f(x1|x2) is approximately the

conditional probability that X1 lies in a small neighborhood (of length h1) of x1,

given that X2 lies in a small neighborhood of x2. A similar interpretation may

be given to f(x2|x1). We can also define the conditional d.f. of X2, given X1 = x1,

by means of

4.2 The 4.1

d.f. ofThe

a Random

Cumulative

Vector

Distribution

and Its Properties

Function 95

f x x ( )

(

)

2 1

F x2 x1 = x x 2 2

x f x x dx ,

2 1 2

2

( )

and similarly for F(x1|x2).

The concepts introduced thus far generalize in a straightforward way for

k > 2. Thus if X = (X1, . . . , Xk) with p.d.f. f(x1, . . . , xk), then we have called

f(x1, . . . , xk) the joint p.d.f. of the r.v.s X1, X2, . . . , Xk. If we sum (integrate)

over t of the variables x1, . . . , xk keeping the remaining s fixed (t + s = k), the

resulting function is the joint p.d.f. of the r.v.s corresponding to the remaining

s variables; that is,

f x1 , , xk

x , , x

( )

(

fi , , i xi , , xi ) = j1 jt

( )

1 s 1 s

f x ,

1 , xk dx j dx j . 1 t

There are

k k k

+ + + =2 2

k

1 2 k 1

such p.d.f.s which are also called marginal p.d.f.s. Also if xi1, . . . , xis are such

that fi1, . . . , it (xi1, . . . , xis) > 0, then the function (of xj1, . . . , xjt) defined by

(

f x1 , , xk )

(

f x j , , x j xi , , xi =

1 t 1 s ) (

fi , , i xi , , xi )

1 s 1 s

is a p.d.f. called the joint conditional p.d.f. of the r.v.s Xj1, . . . , Xjt, given Xi1 =

xi1, , Xjs = xjs, or just given Xi1, . . . , Xis. Again there are 2k 2 joint condi-

tional p.d.f.s involving all k r.v.s X1, . . . , Xk. Conditional distribution func-

tions are defined in a way similar to the one for k = 2. Thus

(

F x j , , x j xi , , xi

1 t 1 t)

f (x , , x x , , x )

j j i i 1 t 1 s

= ( x , , x )( x , , x

j1 jt j1 jt )

x x f x ,

jt

(

j , x j x i , , x i dx j dx j .

j1

1 t 1 s ) 1 t

taken from a discrete distribution and the other taken from a continuous

distribution.

EXAMPLE 1 Let the r.v.s X1, . . . , Xk have the Multinomial distribution with parameters

n and p1, . . . , pk. Also, let s and t be integers such that 1 s, t < k and

s + t = k. Then in the notation employed above, we have:

96 4 Distribution Functions, Probability Densities, and Their Relationship

ii) 1 s

(

fi , , i xi , , xi =

1 s

) n!

xi ! xi ! n r !

x

(

x

pi pi q n r ,

)

i1

1 s

is

1 s

(

q = 1 pi + + pi , r = xi + + xi ;

1 s

) 1 s

that is, the r.v.s Xi1, . . . , Xis and Y = n (Xi1 + + Xis) have the Multinomial

distribution with parameters n and pi1, . . . , pis, q.

(n r )! p j1

x j1

pj t

xj

ii) (

f x j1 , , x jt xi1 , , xi s ) =

x j1 ! x jt ! q

1 ,

q

r = xi1 + + xi s ;

that is, the (joint) conditional distribution of Xj1, . . . , Xjt given Xi1, . . . , Xis is

Multinomial with parameters n r and pj1/q, . . . , pjt/q.

DISCUSSION

i) Clearly,

(X i1 1 s s

) (

= xi , , X i = xi X i + + X i = r = n Y = r = Y = n r ,

1 s

) ( ) ( )

so that

(X i1 s s

) (

= xi , , X i = xi = X i = xi , , X i = xi , Y = n r .

1 1 1 s s

)

Denoting by O the outcome which is the grouping of all n outcomes distinct

from those designated by i1, . . . , is , we have that the probability of O is q, and

the number of its occurrences is Y. Thus, the r.v.s Xi1, . . . , Xis and Y are

distributed as asserted.

ii) We have

) ( f ( x , , x ) ) = f ((x , , x ))

f x , , x ,x , ,x f x, ,x

(

f x j1 , , x jt xi1 , , xi s =

j1

i1

jt i1

is

is 1

i1

k

is

n! 1 k

n! x x

= p1x pkx pi1i1 pi si s q n r

x1! xk ! xi1 ! xi s ! n r ! ( )

p p p p

xi1 xi s x j1 x jt p p q

xi1 xi s x j1 + + x jt

i is j1 jt i is

= 1 1

xi1 ! xi s ! x j1 ! x jt !

xi ! xi ! n r !

1 s ( )

(since n r = n ( x i1 )

+ + xi s = x j1 + + x jt )

=

(n r )! p j1

x j1

pj

t ,

x jt

x j1 ! x jt ! q q

as was to be seen.

EXAMPLE 2 Let the r.v.s X1 and X2 have the Bivariate Normal distribution, and recall that

their (joint) p.d.f. is given by:

4.1 The Cumulative DistributionExercises

Function 97

(

f x1 , x2 = ) 1

2 1 2 1 2

x 2 2

1 x1 1 x2 2 x2 2

exp 1 1

2 + .

(

2 1

2

) 1

1 2 2

We saw that the marginal p.d.f.s f1, f2 are N(1, 21), N(2, 22), respectively;

that is, X1, X2 are also normally distributed. Furthermore, in the process of

proving that f(x1, x2) is a p.d.f., we rewrote it as follows:

x

( )

( )

2 2

x b

(

f x1 , x 2 = ) 1

exp

1

2 12

1 exp

2

2

,

2 1 2 1 2 2

2 2 1

where

2

b = 2 +

1

(

x1 1 . )

Hence

( )= ( )

2

f x1 , x 2 x2 b

(

f x 2 x1 = ) ( )

f1 x1 2 2

1

1 2

exp

2

2

2 2 1

which is the p.d.f. of an N(b, 22(1 2)) r.v. Similarly f(x1|x2) is seen to be the

p.d.f. of an N(b, 21(1 2)) r.v., where

1

b = 1 +

2

(

x2 2 . )

Exercises

4.2.1 Refer to Exercise 3.2.17 in Chapter 3 and:

i) Find the marginal p.d.f.s of the r.v.s Xj, j = 1, , 6;

ii) Calculate the probability that X1 5.

4.2.2 Refer to Exercise 3.2.18 in Chapter 3 and determine:

ii) The marginal p.d.f. of each one of X1, X2, X3;

ii) The conditional p.d.f. of X1, X2, given X3; X1, X3, given X2; X2, X3, given

X1 ;

98 4 Distribution Functions, Probability Densities, and Their Relationship

iii) The conditional p.d.f. of X1, given X2, X3; X2, given X3, X1; X3, given X1,

X2 .

If n = 20, provide expressions for the following probabilities:

iv) P(3X1 + X2 5);

v) P(X1 < X2 < X3);

vi) P(X1 + X2 = 10|X3 = 5);

vii) P(3 X1 10|X2 = X3);

viii) P(X1 < 3X2|X1 > X3).

4.2.3 Let X, Y be r.v.s jointly distributed with p.d.f. f given by f(x, y) = 2/c2

if 0 x y, 0 y c and 0 otherwise.

i) Determine the constant c;

ii) Find the marginal p.d.f.s of X and Y;

iii) Find the conditional p.d.f. of X, given Y, and the conditional p.d.f. of Y,

given X;

iv) Calculate the probability that X 1.

4.2.4 Let the r.v.s X, Y be jointly distributed with p.d.f. f given by f(x, y) =

exy I(0,)(0,) (x, y). Compute the following probabilities:

i) P(X x);

ii) P(Y y);

iii) P(X < Y);

iv) P(X + Y 3).

4.2.5 If the joint p.d.f. f of the r.v.s Xj, j = 1, 2, 3, is given by

( )

f x1 , x 2 , x3 = c 3 e

(

c x1 + x2 + x 3 )

( )

I A x1 , x 2 , x3 ,

where

( ) (

A = 0, 0, 0, , ) ( )

i) Determine the constant c;

ii) Find the marginal p.d.f. of each one of the r.v.s Xj, j = 1, 2, 3;

iii) Find the conditional (joint) p.d.f. of X1, X2, given X3, and the conditional

p.d.f. of X1, given X2, X3;

iv) Find the conditional d.f.s corresponding to the conditional p.d.f.s in (iii).

4.2.6 Consider the function given below:

yx e y

( )

f x y = x!

0,

, x = 0, 1, ; y 0

otherwise.

4.3

4.1 Quantiles

The Cumulative

and Modes

Distribution

of a Distribution

Function 99

i) Show that for each fixed y, f(|y) is a p.d.f., the conditional p.d.f. of an r.v.

X, given that another r.v. Y equals y;

ii) If the marginal p.d.f. of Y is Negative Exponential with parameter = 1,

what is the joint p.d.f. of X, Y?

iii) Show that the marginal p.d.f. of X is given by f(x) = ( 12 )x+1 IA(x), where

A = {0, 1, 2, . . . }.

4.2.7 Let Y be an r.v. distributed as P() and suppose that the conditional

distribution of the r.v. X, given Y = n, is B(n, p). Determine the p.d.f. of X and

the conditional p.d.f. of Y, given X = x.

4.2.8 Consider the function f defined as follows:

x 12 + x 22

( )

f x1 , x 2 =

1

2

exp

2 4e

+

1 3 3

(

x1 x 2 I [ 1, 1][ 1, 1] x1 , x 2 )

and show that:

i) f is a non-Normal Bivariate p.d.f.

ii) Both marginal p.d.f.s

( ) ( )

f1 x1 = f x1 , x 2 dx 2

and

( ) ( )

f2 x 2 = f x1 , x 2 dx1

Let X be an r.v. with d.f. F and consider a number p such that 0 < p < 1. A

pth quantile of the r.v. X, or of its d.f. F, is a number denoted by xp and having

the following property: P(X xp) p and P(X xp) 1 p. For p = 0.25 we

get a quartile of X, or its d.f., and for p = 0.5 we get a median of X, or its

d.f. For illustrative purposes, consider the following simple examples.

EXAMPLE 3 Let X be an r.v. distributed as U(0, 1) and let p = 0.10, 0.20, 0.30, 0.40, 0.50,

0.60, 0.70, 0.80 and 0.90. Determine the respective x0.10, x0.20, x0.30, x0.40, x0.50, x0.60,

x0.70, x0.80, and x0.90.

Since for 0 x 1, F(x) = x, we get: x0.10 = 0.10, x0.20 = 0.20, x0.30 = 0.30,

x0.40 = 0.40, x0.50 = 0.50, x0.60 = 0.60, x0.70 = 0.70, x0.80 = 0.80, and x0.90 = 0.90.

EXAMPLE 4 Let X be an r.v. distributed as N(0, 1) and let p = 0.10, 0.20, 0.30, 0.40, 0.50,

0.60, 0.70, 0.80 and 0.90. Determine the respective x0.10, x0.20, x0.30, x0.40, x0.50, x0.60,

x0.70, x0.80, and x0.90.

100 4 Distribution Functions, Probability Densities, and Their Relationship

Typical cases:

F(x) F(x)

1

p

p

x [ ] x

xp 0 0

xp

(a) (b)

F(x)

x

0 xp

(c)

F(x) F(x)

1

1

p

p

x [ ] x

0 xp 0

xp

(d) (e)

Figure 4.3 Observe that the figures demonstrate that, as defined, xp need not be unique.

and symmetry, we find: x0.10 = 1.282, x0.20 = 0.842, x0.30 = 0.524, x0.40 = 0.253,

x0.50 = 0, x0.60 = 0.253, x0.70 = 0.524, x0.80 = 0.842, and x0.90 = 1.282.

Knowledge of quantiles xp for several values of p provides an indication as

to how the unit probability mass is distributed over the real line. In Fig. 4.3

various cases are demonstrated for determining graphically the pth quantile of

a d.f.

Let X be an r.v. with a p.d.f. f. Then a mode of f, if it exists, is any number

which maximizes f(x). In case f is a p.d.f. which is twice differentiable, a mode

can be found by differentiation. This process breaks down in the discrete cases.

The following theorems answer the question for two important discrete cases.

4.3

4.1 Quantiles

The Cumulative

and Modes

Distribution

of a Distribution

Function 101

n

()

f x = px qn x ,

x

0 < p < 1, q = 1 p, x = 0, 1, , n.

Consider the number (n + 1)p and set m = [(n + 1)p], where [y] denotes the

largest integer which is y. Then if (n + 1)p is not an integer, f(x) has a unique

mode at x = m. If (n + 1)p is an integer, then f(x) has two modes obtained for

x = m and x = m 1.

PROOF For x 1, we have

n x n x

p q

f x

=

() x

(

f x1

)

n x 1 n x +1

p q

x 1

n!

p x q n x

=

(

x! n x ! ) =

nx +1 p

.

n! x q

p x 1 q n x +1

( )(

x 1! nx +1! )

That is,

f x() =

nx +1 p

.

(

f x1 ) x q

Thus if (n + 1)p is not an integer, f(x) keeps increasing for x m and then

decreases so the maximum occurs at x = m. If (n + 1)p is an integer, then the

maximum occurs at x = (n + 1)p, where f(x) = f(x 1) (from above calcula-

tions). Thus

(

x = n+1 p1 )

is a second point which gives the maximum value. "

THEOREM 6 Let X be P(); that is,

x

()

f x = e

x!

, x = 0, 1, 2, , > 0.

then f(x) has two modes obtained for x = and x = 1.

PROOF For x 1, we have

102 4 Distribution Functions, Probability Densities, and Their Relationship

f x () =

( )

e x x!

=

.

(

f x1 ) e

[ (x 1)!]

x 1 x

Hence f(x) > f(x 1) if and only if > x. Thus if is not an integer, f(x) keeps

increasing for x [] and then decreases. Then the maximum of f(x) occurs

at x = []. If is an integer, then the maximum occurs at x = . But in this case

f(x) = f(x 1) which implies that x = 1 is a second point which gives

the maximum value to the p.d.f. "

Exercises

4.3.1 Determine the pth quantile xp for each one of the p.d.f.s given in

Exercises 3.2.1315, 3.3.1316 (Exercise 3.2.14 for = 14 ) in Chapter 3 if p =

0.75, 0.50.

4.3.2 Let X be an r.v. with p.d.f. f symmetric about a constant c (that is,

f(c x) = f(c + x) for all x ! ). Then show that c is a median of f.

4.3.3 Draw four graphstwo each for B(n, p) and P()which represent

the possible occurrences for modes of the distributions B(n, p) and P().

4.3.4 Consider the same p.d.f.s mentioned in Exercise 4.3.1 from the point

of view of a mode.

In this section, a rigorous justification of Statements 1 and 2 made in Section

4.1 will be presented. For this purpose, some preliminary concepts and results

are needed and will be also discussed.

DEFINITION 1 A set G in ! is called open if for every x in G there exists an open interval

containing x and contained in G. Without loss of generality, such intervals may

be taken to be centered at x.

It follows from this definition that an open interval is an open set, the

entire real line ! is an open set, and so is the empty set (in a vacuous manner).

LEMMA 1 Every open set in ! is measurable.

PROOF Let G be an open set in !, and for each x G, consider an open

interval centered at x and contained in G. Clearly, the union over x, as x varies

in G, of such intervals is equal to G. The same is true if we consider only those

intervals corresponding to all rationals x in G. These intervals are countably

many and each one of them is measurable; then so is their union. "

4.1

4.4*The

Justification

CumulativeofDistribution

StatementsFunction

1 and 2 103

DEFINITION 2 A set G in ! m, m 1, is called open if for every x in G there exists an open cube

in ! m containing x and contained in G; by the term open cube we mean the

Cartesian product of m open intervals of equal length. Without loss of gener-

ality, such cubes may be taken to be centered at x.

LEMMA 2 Every open set in ! n is measurable.

PROOF It is analogous to that of Lemma 1. Indeed, let G be an open set in

! m, and for each x G, consider an open cube centered at x and contained in

G. The union over x, as x varies in G, of such cubes clearly is equal to G. The

same is true if we restrict ourselves to xs in G whose m coordinates are

rationals. Then the resulting cubes are countably many, and therefore their

union is measurable, since so is each cube. "

DEFINITION 3 Recall that a function g: S ! ! is said to be continuous at x0 S if for

every > 0 there exists a = (, x0) > 0 such that |x x0| < implies |g(x) g(x0)|

< . The function g is continuous in S if it is continuous for every x S.

It follows from the concept of continuity that 0 implies 0.

LEMMA 3 Let g: ! ! be continuous. Then g is measurable.

PROOF By Theorem 5 in Chapter 1 it suffices to show that g1(G) are meas-

urable sets for all open intevals G in !. Set B = g1(G). Thus if B = , the

assertion is valid, so let B and let x0 be an arbitrary point of B, so that g(x0)

G. Continuity of g at x0 implies that for every > 0 there exists = (, x0)

> 0 such that |x x0| < implies |g(x) g(x0)| < . Equivalently, x (x0 , x0

+ ) implies g(x) (g(x0) , g(x0) + ). Since g(x0) G and G is open, by

choosing sufficiently small, we can make so small that (g(x0) , g(x0) + )

is contained in G. Thus, for such a choice of and , x (x0 , x0 + ) implies

that (g(x0) , g(x0) + ) G. But B(= g1(G)) is the set of all x ! for which

g(x) G. As all x (x0 , x0 + ) have this property, it follows that (x0 , x0

+ ) B. Since x0 is arbitrary in B, it follows that B is open. Then by Lemma

1, it is measurable. "

The concept of continuity generalizes, of course, to Euclidean spaces of

higher dimensions, and then a result analogous to the one in Lemma 3 also

holds true.

every > 0 there exists a = (, x0) > 0 such that ||x x0|| < implies ||g(x)

g(x0)|| < . The function g is continuous in S if it is continuous for every x S.

Here ||x|| stands for the usual norm in ! k; i.e., for x = (x1, . . . , xk), ||x|| =

(ik=1 xi2 ) , and similarly for the other quantities.

12

0.

LEMMA 4 Let g: ! k ! m be continuous. Then g is measurable.

104 4 Distribution Functions, Probability Densities, and Their Relationship

PROOF The proof is similar to that of Lemma 3. The details are presented

here for the sake of completeness. Once again, it suffices to show that g1(G)

are measurable sets for all open cubes G in !m. Set B = g1(G). If B = the

assertion is true, and therefore suppose that B and let x0 be an arbitrary

point of B. Continuity of g at x0 implies that for every > 0 there exists a =

(, x0) > 0 such that ||x x0|| < implies ||g(x) g(x0)|| < ; equivalently, x

S(x0, ) implies g(x) S(g(x0), ), where S(c, r) stands for the open sphere with

center c and radius r. Since g(x0) G and G is open, we can choose so small

that the corresponding is sufficiently small to imply that g(x) S(g(x0), ).

Thus, for such a choice of and , x S(x0, ) implies that g(x) S(g(x0), ).

Since B(= g1(G)) is the set of all x ! k for which g(x) G, and x S(x0, )

implies that g(x) S(g(x0), ), it follows that S(x0, ) B. At this point, observe

that it is clear that there is a cube containing x0 and contained in S(x0, ); call

it C(x0, ). Then C(x0, ) B, and therefore B is open. By Lemma 2, it is also

measurable. "

We may now proceed with the justification of Statement 1.

THEOREM 7 Let X : (S, A) (! k, B k) be a random vector, and let g : (! k, B k) (! m, Bm)

be measurable. Then g(X): (S, A) (! m, Bm) and is a random vector. (That

is, measurable functions of random vectors are random vectors.)

PROOF To prove that [g(X)]1(B) A if B B m, we have

1

1 1 1

1 where B1 = g 1 B B k

1

by the measurability of g. Also, X (B1) A since X is measurable. The proof

is completed. "

To this theorem, we have the following

COROLLARY Let X be as above and g be continuous. Then g(X) is a random vector. (That

is, continuous functions of random vectors are random vectors.)

PROOF The continuity of g implies its measurability by Lemma 3, and there-

fore the theorem applies and gives the result. "

DEFINITION 5 For j = 1, . . . , k, the jth projection function gj is defined by: gj: ! k ! and

gj(x) = gj(x1, . . . , xk) = xj.

It so happens that projection functions are continuous; that is,

LEMMA 5 The coordinate functions gj, j = 1, . . . , k, as defined above, are continuous.

PROOF For an arbitrary point x0 in ! K, consider x ! K such that ||x x0|| <

for some > 0. This is equivalent to ||x x0||2 < 2 or j =1 ( x j x ) < 2 which

k 2

0j

implies that (xj x0j)2 < 2 for j = 1, . . . , k, or |xj x0j| < , j = 1, . . . , k. This last

expression is equivalent to |gj(x) gj(x0)| < , j = 1, . . . , k. Thus the definition

of continuity of gj is satisfied here for = . "

Now consider a k-dimensional function X defined on the sample space S.

Then X may be written as X = (X1, . . . , Xk), where Xj, j = 1, . . . , k are real-

valued functions. The question then arises as to how X and Xj, j = 1, . . . , k are

4.1 The Cumulative DistributionExercises

Function 105

related from a measurability point of view. To this effect, we have the follow-

ing result.

Let X = (X1, . . . , Xk) : (S, A) (! k, B k). Then X is an r. vector if and only if

THEOREM 8 Xj, j = 1, . . . , k are r.v.s.

PROOF Suppose X is an r. vector and let gj, j = 1, . . . , k be the coordinate

functions defined on ! k. Then gjs are continuous by Lemma 5 and therefore

measurable by Lemma 4. Then for each j = 1, . . . , k, gj(X) = gj(X1, . . . , Xk) =

Xj is measurable and hence an r.v.

Next, assume that Xj, j = 1, . . . , k are r.v.s. To show that X is an r. vector,

by special case 3 in Section 2 of Chapter 1, it suffices to show that X1(B) A

for each B = (, x1] (, xk], x1, . . . , xk !. Indeed,

k

X 1 ( B) = (X B) = ( X j ( , x j ], j = 1, , k ) = I X j1 (( , x j ]) A .

j =1

Exercises

4.4.1 If X and Y are functions defined on the sample space S into the real line

!, show that:

r Q

4.4.2 Use Exercise 4.4.1 in order to show that, if X and Y are r.v.s, then so

is the function X + Y.

4.4.3

ii) If X is an r.v., then show that so is the function X.

ii) Use part (i) and Exercise 4.4.2 to show that, if X and Y are r.v.s, then so is

the function X Y.

4.4.4

ii) If X is an r.v., then show that so is the function X 2.

1

ii) Use the identity: XY = 2 (X + Y)2 12 (X 2 + Y 2) in conjunction with part (i)

and Exercises 4.4.2 and 4.4.3(ii) to show that, if X and Y are r.v.s, then so

is the function XY.

4.4.5

ii) If X is an r.v., then show that so is the function 1

X

, provided X 0.

ii) Use part (i) in conjunction with Exercise 4.4.4(ii) to show that, if X and Y

are r.v.s, then so is the function XY , provided Y 0.

106 5 Moments of Random VariablesSome Moment and Probability Inequalities

Chapter 5

Moment and Probability Inequalities

In the definitions to be given shortly, the following remark will prove useful.

REMARK 1 We say that the (infinite) series xh(x), where x = (x1, . . . , xk)

varies over a discrete set in !k, k 1, converges absolutely if x|h(x)| < . Also

( )

we say that the integral h x1 , , xk dx1 dxk converges absolutely

if

assumed that they converge absolutely. In this case, we say that the moments to

be defined below exist.

Let X = (X1, . . . , Xk) be an r. vector with p.d.f. f and consider the (measurable)

function g: ! k !, so that g(X) = g(X1, . . . , Xk) is an r.v. Then we give the

DEFINITION 1 iii) For n = 1, 2, . . . , the nth moment of g(X) is denoted by E[g(X)]n and is

defined by:

[ ( )] ( ) (

)

n

g x f x , x = x1 , , xk

[ ( )]

n

E g X = x

[(

g x ,

)] ( )

n

1 , xk f x1 , , xk dx1 dxk .

For n = 1, we get

g x f x ()()

[ ( )]

E g X = x

( )( )

g x1 , , xk f x1 , , xk dx1 dxk

106

5.1 Moments of Random Variables 107

and call it the mathematical expectation or mean value or just mean of g(X).

Another notation for E[g(X)] which is often used is g(X), or [g(X)], or just

, if no confusion is possible.

iii) For r > 0, the rth absolute moment of g(X) is denoted by E|g(X)|r and is

defined by:

() () ( )

r

g x f x , x = x1 , , xk

( )

r

Eg X = x

g x ,

( ) ( )

r

1 , xk f x1 , , xk dx1 dxk .

iii) For an arbitrary constant c, and n and r as above, the nth moment and rth

absolute moment of g(X) about c are denoted by E[g(X) c]n, E|g(X) c|r,

respectively, and are defined as follows:

[() ] () (

)

n

g x c f x , x = x1 , , xk

[( ) ]

n

E g X c = x

[( ) ] (

g x ,

)

n

1 , xk c f x1 , , xk dx1 dxk ,

and

() () ( )

r

g x c f x , x = x1 , , xk

( )

r

Eg X c = x

g x ,

( ) ( )

r

1 , xk c f x1 , , xk dx1 dxk .

For c = E[g(X)], the moments are called central moments. The 2nd central

moment of g(X), that is,

{( ) [ ( )]}

2

E g X E g X

[ g( x) Eg(X )] f ( x),

( )

2

x = x1 , , xk

= x

g x ,

[( ) ( )] f ( x , , x )dx

2

1 , xk Eg X 1 k 1 dxk

is called the variance of g(X) and is also denoted by 2[g(X)], or g2( X ), or just

2, if no confusion is possible. The quantity + 2 [ g(X )] = [ g(X )] is called the

standard deviation (s.d.) of g(X) and is also denoted by g(X), or just , if no

confusion is possible. The variance of an r.v. is referred to as the moment of

inertia in Mechanics.

1. Let g(X1, . . . , Xk) = X 1 X k , where nj 0 are integers. Then

n 1 n k

1 k

108 5 Moments of Random VariablesSome Moment and Probability Inequalities

()

x nj f x =

x

(

x nj f x1 , , xk )

( )

E X n

= ( x , ,x )

1 k

j

xn f x ,

( )

j 1 , xk dx1 dxk

( )

x nj f j x j

x

= j

j j j ( )

x n f x dx

j

which is the nth moment of the r.v. Xj. Thus the nth moment of an r.v. X with

p.d.f. f is

x n f x ()

( )

E Xn

= x

()

x n f x dx.

For n = 1, we get

xf x

()

( )

E X = x

()

xf x dx

quantity is also denoted by X or (X) or just when no confusion is possible.

The quantity X can be interpreted as follows: It follows from the defini-

tion that if X is a discrete uniform r.v., then X is just the arithmetic average of

the possible outcomes of X. Also, if one recalls from physics or elementary

calculus the definition of center of gravity and its physical interpretation as the

point of balance of the distributed mass, the interpretation of X as the mean

or expected value of the random variable is the natural one, provided the

probability distribution of X is interpreted as the unit mass distribution.

REMARK 2 In Definition 1, suppose X is a continuous r.v. Then E[g(X)] =

g( x)f ( x)dx. On the other hand, from the last expression above, E(X) =

More specifically, in the definition of E[g(X)], one would expect to use the

p.d.f. of g(X) rather than that of X. Actually, the definition of E[g(X)], as

given, is correct and its justification is roughly as follows: Consider E[g(x)] =

g( x)f ( x)dx and set y = g(x). Suppose that g is differentiable and has an

inverse g1, and that some further conditions are met. Then

1 1 d

5.1 Moments of Random Variables 109

Therefore the last integral above is equal to yfY ( y)dy, which is consonant

2. For g as above, that is, g(X1, . . . , Xk) = X 1n X kn and n1 = = nj1 = 1 k

( ) () ( )

n

x j EX j f x , x = x1 , , xk

( )

n

E X j EX j = x

(

x EX n f x ,

j j ) (

1 , xk dx1 dxk )

( ) ( )

n

x j EX j f j x j

= xj

(

x EX n f x dx

j j j j j ) ( )

which is the nth central moment of the r.v. Xj (or the nth moment of Xj about its

mean).

Thus the nth central moment of an r.v. X with p.d.f. f and mean is

( ) f (x) = (x ) f (x)

n n

x EX

( ) ( )

n n

E X EX =E X = x x

x EX ( ) f (x)dx = (x ) f (x)dx.

n n

or just 2 when no confusion is possible, and is called the variance of X. Its

positive square root X or (X) or just is called the standard deviation (s.d.)

of X.

As in the case of X, X2 has a physical interpretation also. Its definition

corresponds to that of the second moment, or moment of inertia. One recalls

that a large moment of inertia means the mass of the body is spread widely

about its center of gravity. Likewise a large variance corresponds to a probabil-

ity distribution which is not well concentrated about its mean value.

3. For g(X1, . . . , Xk) = (X1 EX1)n (Xk EXk)n , the quantity 1 k

(

E X1 EX1 X k EX k ) ( )

n n 1 k

is the (n1, . . . , nk)-central joint moment of X1, . . . , Xk or the (n1, . . . , nk)-joint

moment of X1, . . . , Xk about their means.

4. For g(X1, . . . , Xk) = Xj(Xj 1) (Xj n + 1), j = 1, . . . , k, the quantity

x j x j 1 x j n + 1 f j x j ( ) ( ) ( )

[ ( )

x

E Xj Xj 1 Xj n+1 = ( )] j

x x 1 x n + 1 f x dx

j j j j j j ( ) ( ) ( )

110 5 Moments of Random VariablesSome Moment and Probability Inequalities

is the nth factorial moment of the r.v. Xj. Thus the nth factorial moment of an

r.v. X with p.d.f. f is

(

x x 1 x n + 1 f x ) ( )()

[ ( )

(

E X X 1 X n+1 = x )]

( ) ( )()

x x 1 x n + 1 f x dx.

5.1.2 Basic Properties of the Expectation of an R.V.

From the very definition of E[g(X)], the following properties are immediate.

(E1) E(c) = c, where c is a constant.

(E2) E[cg(X)] = cE[g(X)], and, in particular, E(cX) = cE(X) if X is an

r.v.

(E3) E[g(X) + d] = E[g(X)] + d, where d is a constant. In particular,

E(X + d) = E(X) + d if X is an r.v.

(E4) Combining (E2) and (E3), we get E[cg(X) + d] = cE[g(X)] + d,

and, in particular, E(cX + d) = cE(X) + d if X is an r.v.

[ ( )] [ ( )]

(E4) E nj = 1 cj g j X = nj = 1 cj E g j X .

In fact, for example, in the continuous case, we have

n n

( ) ( ) ( )

E c j g j X = c j g j x1 , , xk f x1 , , xk dx1 dxk

j =1

j =1

n

( )( )

= c j g j x1 , , xk f x1 , , xk dx1 dxk

j =1

[ ( )]

n

= cjE gj X .

j =1

(E4) E ( jn= 1 c j X j ) = jn= 1 c j E ( X j ).

(E5) If X 0, then E(X) 0.

Consequently, by means of (E5) and (E4), we get that

(E5) If X Y, then E(X) E(Y), where X and Y are r.v.s (with finite

expectations).

(E6) |E[g(X)]| E|g(X)|.

(E7) If E|X|r < for some r > 0, where X is an r.v., then E|X|r< for

all 0 < r < r.

This is a consequence of the obvious inequality |X|r 1 + |X|r and (E5).

Furthermore, since of n = 1, 2, . . . , we have |Xn| = |X|n, by means of (E6),

it follows that

5.1 Moments of Random Exercises

Variables 111

(E7) If E(Xn) exists (that is, E|X|n < ) for some n = 2, 3, . . . , then E(Xn)

also exists for all n = 1, 2, . . . with n < n.

Regarding the variance, the following properties are easily established by

means of the definition of the variance.

(V1) 2(c) = 0, where c is a constant.

(V2) 2[cg(X)] = c2 2[g(X)], and, in particular, 2(cX) = c2 2(X), if X is

an r.v.

(V3) 2[g(X) + d] = 2[g(X)], where d is a constant. In particular,

2(X + d) = 2(X), if X is an r.v.

In fact,

[( ) ] {[ ( ) ] [ ( ) ]}

2

2 g X +d = E g X +d E g X +d

2

2

and, in particular, 2(cX + d) = c2 2(X), if X is an r.v.

(V5) 2[g(X)] = E[g(X)]2 [Eg(X)]2, and, in particular,

(V5) 2(X) = E(X 2) (EX)2, if X is an r.v.

In fact,

2 2 2

2 g X = E g X Eg X

2 2 2 2 2

the equality before the last one being true because of (E4).

(V6) 2(X) = E[X(X 1)] + EX (EX)2, if X is an r.v., as is easily seen.

This formula is especially useful in calculating the variance of a

discrete r.v., as is seen below.

Exercises

5.1.1 Verify the details of properties (E1)(E7).

5.1.2 Verify the details of properties (V1)(V5).

5.1.3 For r < r, show that |X|r 1 + |X|r and conclude that if E|X|r < , then

E|X|r for all 0 < r < r.

112 5 Moments of Random VariablesSome Moment and Probability Inequalities

5.1.4 Verify the equality (E[ g( X )] =) g( x) fX ( x)dx = yfY ( y)dy for the

case that X N(0, 1) and Y = g(X) = X2.

5.1.5 For any event A, consider the r.v. X = IA, the indicator of A defined by

IA(s) = 1 for s A and IA(s) = 0 for s Ac, and calculate EXr, r > 0, and also

2(X).

5.1.6 Let X be an r.v. such that

( ) (

P X = c = P X = c = .

2

1

)

Calculate EX, 2(X) and show that

( ).

2 X

(

P X EX c = ) c2

5.1.7 Let X be an r.v. with finite EX.

ii) For any constant c, show that E(X c)2 = E(X EX)2 + (EX c)2;

ii) Use part (i) to conclude that E(X c)2 is minimum for c = EX.

5.1.8 Let X be an r.v. such that EX4 < . Then show that

ii) E(X EX)3 = EX3 3(EX)(EX)2 + 2(EX)3;

ii) E(X EX)4 = EX4 4(EX)(EX3) + 6(EX)2(EX2) 3(EX)4.

5.1.9 If EX4 < , show that:

[ ( )] [ ( )( )]

E X X 1 = EX 2 EX ; E X X 1 X 2 = EX 3 3EX 2 + 2EX ;

4 3 2

6EX .

the factorial moments E[X(X 1)], E[X(X 1)(X 2)], E[X(X 1)(X 2)

(X 3)].)

5.1.10 Let X be the r.v. denoting the number of claims filed by a policy-

holder of an insurance company over a specified period of time. On the basis

of an extensive study of the claim records, it may be assumed that the distribu-

tion of X is as follows:

x 0 1 2 3 4 5 6

iii) What premium should the company charge in order to break even?

iii) What should be the premium charged if the company is to expect to come

ahead by $M for administrative expenses and profit?

5.1 Moments of Random Exercises

Variables 113

5.1.11 A roulette wheel has 38 slots of which 18 are red, 18 black, and 2

green.

iii) Suppose a gambler is placing a bet of $M on red. What is the gamblers

expected gain or loss and what is the standard deviation?

iii) If the same bet of $M is placed on green and if $kM is the amount

the gambler wins, calculate the expected gain or loss and the standard

deviation.

iii) For what value of k do the two expectations in parts (i) and (ii) coincide?

iv) Does this value of k depend on M?

iv) How do the respective standard deviations compare?

5.1.12 Let X be an r.v. such that P(X = j) = ( 12 )j, j = 1, 2, . . . .

ii) Compute EX, E[X(X 1)];

ii) Use (i) in order to compute 2(X).

5.1.13 If X is an r.v. distributed as U(, ), show that

( )

2

+

EX =

2

, X2

( ) =

12

.

5.1.14 Let the r.v. X be distributed as U(, ). Calculate EXn for any positive

integer n.

5.1.15 Let X be an r.v. with p.d.f. f symmetric about a constant c (that is,

f(c x) = f(c + x) for every x).

ii) Then if EX exists, show that EX = c;

ii) If c = 0 and EX2n+1 exists, show that EX2n+1 = 0 (that is, those moments of X

of odd order which exist are all equal to zero).

5.1.16 Refer to Exercise 3.3.13(iv) in Chapter 3 and find the EX for those s

for which this expectation exists, where X is an r.v. having the distribution in

question.

5.1.17 Let X be an r.v. with p.d.f. given by

x

()

f x = ()

I c ,c x .

c2 ( )

5.1.18 Let X be an r.v. with finite expectation and d.f. F.

ii) Show that

[ ( )] ()

0

EX = 1 F x dx F x dx;

0

114 5 Moments of Random VariablesSome Moment and Probability Inequalities

ii) Use the interpretation of the definite integral as an area in order to give a

geometric interpretation of EX.

5.1.19 Let X be an r.v. of the continuous type with finite EX and p.d.f. f.

ii) If m is a median of f and c is any constant, show that

( )()

c

E X c = E X m + 2 c x f x dx;

m

ii) Utilize (i) in order to conclude that E|X c| is minimized for c = m. (Hint:

Consider the two cases that c m and c < m, and in each one split the

integral from to c and c to in order to remove the absolute value.

Then the fact that f ( x)dx = m f ( x)dx = 21 and simple manipulations

m

prove part (i). For part (ii), observe that m (c x) f ( x)dx 0 whether c m

c

or c < m.)

5.1.20 If the r.v. X is distributed according to the Weibull distribution (see

Exercise 4.1.15 in Chapter 4), then:

1 1 2 2

ii) Show that EX = 1 + , EX 2 = 1 + , so that

2 1

( )

2 X = 1 + 2 1 + 2 ,

where recall that the Gamma function is defined by = 0 t 1 e t dt , ()

> 0;

ii) Determine the numerical values of EX and 2(X) for = 1 and = 12 ,

= 1 and = 2.

5.2.1 Discrete Case

1. Let X be B(n, p). Then E(X) = np, 2(X) = npq. In fact,

n

n n

( )

E X = x p x q n x = x

n!

px qn x

x = 0 x x =1 x! n x ! ( )

=

n

( )

n n1 !

px qn x

( x 1)!(n x)!

x =1

= np

n

(n 1)! p q

( )( ) x 1 n 1 x 1

( ) [( ) ( )]

x

x =11 ! n 1 x 1 !

(n 1)! p q( ) = np p + q

n 1

( )

n 1

= np

n 1 x

x

= np.

x![( n 1) x]!

x= 0

5.2 Expectations

5.1 Moments

and Variances

of Random

of Some

Variables

R.V.s 115

Next,

[ (

E X X 1 )]

n

= x x1 ( ) x! nn! x ! p q x n x

x= 0 ( )

n n( n 1)( n 2)!

= x( x 1)

( 2 x2 )(

n 2 x2 )

p p q

x=2 x( x 1)( x 2)![( n 2) ( x 2)]!

= n( n 1) p 2

n

(n 2)! p q

( )( ) x2 n 2 x2

( x 2)![(n 2) ( x 2)]!

x=2

= n( n 1) p 2 (n 2)! p q( )

n 2

x n 2 x

x![( n 2) x]!

x= 0

= n( n 1) p ( p + q) = n( n 1) p .

2 n 2 2

That is,

[ (

E X X 1 = n n 1 p2 . )] ( )

Hence, by (V6),

( ) [ ( )] ( ) ( )

2

2 X = E X X 1 + EX EX = n n 1 p 2 + np n 2 p 2

2 2 2

= n p np + np n p = np 1 p = npq. 2 2

( )

2. Let X be P(). Then E(X) = 2(X) = . In fact,

x

x

x 1

( )

E X = xe = xe = e

x= 0 x! x =1 x x1 ! (

x =1 x 1 ! ) ( )

x

= e = e e = .

x = 0 x!

Next,

[ ( )] x

E X X 1 = x x 1 e

x= 0

( ) x!

x

x

= x x 1 e ( ) = 2 e = 2 .

x=2 (

x x1 x 2 ! )( x = 0 x! )

Hence EX2 = 2 + , so that, 2(X) = 2 + 2 = .

REMARK 3 One can also prove that the nth factorial moment of X is n; that

is, E[X(X 1) (X n + 1)] = n.

116 5 Moments of Random VariablesSome Moment and Probability Inequalities

1. Let X be N(0, 1). Then

( ) 2( (n)!) ,

2n !

( )

E X 2 n+1 = 0, E X 2 n = n

n 0.

In particular, then

( )

E X = 0, ( )

2 X = E X2 = ( ) 2

2 1!

= 1.

In fact,

( ) 1

2

E X 2 n+1 = x 2 n+1 e x 2

dx.

2

But

0

x dx = x 2n +1e x 2 dx + x 2n +1e x 2 dx

2 n +1 x 2 2 2 2

e

0

0

= y 2n +1e y 2 dy + x 2n +1e x 2 dx

2 2

0

= x 2n +1e x 2 dx + x 2n +1e x 2 dx = 0.

2 2

0 0

x dx = 2 x 2 n e x

2 2

2n

ex 2 2

dx,

0

x 2n e x 2 dx = x 2n 1de x

2 2

2

0 0

( )

2 2

= x 2n 1e x 2

+ 2n 1 x 2n 2 e x 2 dx

0 0

( )

2

= 2n 1 x 2n 2 e x 2 dx,

0

( )

m2 n = 2 n 1 m2 n 2 , and similarly,

m 2 n 2 = (2 n 3)m 2 n 4

M

m2 = 1 m0

(

m0 = 1 since m0 = E X 0 = E 1 = 1 . ( ) () )

Multiplying them out, we obtain

5.2 Expectations

5.1 Moments

and Variances

of Random

of Some

Variables

R.V.s 117

( )( )

m2 n = 2 n 1 2 n 3 1

1 2 ( 2 n 3)( 2 n 2)( 2 n 1)( 2 n) (2 n)!

= =

2 ( 2 n 2)( 2 n) (2 1) [2( n 1)](2 n)

=

(2 n)! =

(2 n)! .

2 [1 ( n 1) n] 2 ( n!)

n n

X 2 X

E = 0, = 1.

But

X 1

E

= EX .

( )

Hence

1

( )

E X = 0,

X

2

1 2

= 2 X ( )

and then

1 2

2

( )

X = 1,

so that 2(X) = 2.

2. Let X be Gamma with parameters and . Then E(X) = and

2(X) = 2. In fact,

( ) 1 1

( ) ( )

EX = xx 1 e x dx = x e x dx

0

0

x e x x 1 e x dx

( ) 0

= x de x =

0

( )

0

( )

= x 1 e x dx = .

0

118 5 Moments of Random VariablesSome Moment and Probability Inequalities

Next,

( ) (1) ( )

E X2 = 0 x +1 e x dx = 2 + 1

and hence

( ) ( )

2 X = 2 + 1 2 2 = 2 + 1 = 2 . ( )

REMARK 5

ii) If X is r2 , that is, if = r/2, = 2, we get E(X) = r, 2(X) = 2r.

ii) If X is Negative Exponential, that is, if = 1, = 1/, we get E(X) = 1/,

2(X) = 1/2.

3. Let X be Cauchy. Then E(Xn) does not exist for any n 1. For example,

for n = 1, we get

x dx

I= .

( )

2

2 + x

1 x dx

1 1 d x

2

( )

1 + x 2 2 1 + x 2

I= =

1 1 d 1+ x

2

( 1 ) ( )

=

2 1+ x 2

=

2

log 1 + x 2

=

1

2

(

, )

which is an indeterminate form. Thus the Cauchy distribution is an example of

a distribution without a mean.

REMARK 6 In somewhat advanced mathematics courses, one encounters

sometimes the so-called Cauchy Principal Value Integral. This coincides with

the improper Riemann integral when the latter exists, and it often exists even

if the Riemann integral does not. It is an improper integral in which the limits

are taken symmetrically. As an example, for = 1, = 0, we have, in terms of

the principal value integral,

I * = lim

A

1 A xdx

A 1+ x 2

=

1

lim log 1 + x 2

2 A

( ) A

A

=

1

[ ( )

lim log 1 + A 2 log 1 + A 2 = 0.

2 A

( )]

5.1 Moments of Random Exercises

Variables 119

Thus the mean of the Cauchy exists in terms of principal value, but not in the

sense of our definition which requires absolute convergence of the improper

Riemann integral involved.

Exercises

5.2.1 If X is an r.v. distributed as B(n, p), calculate the kth factorial moment

E[X(X 1) (X k + 1)].

5.2.2 An honest coin is tossed independently n times and let X be the r.v.

denoting the number of Hs that occur.

iii) Calculate E(X/n), 2(X/n);

iii) If n = 100, find a lower bound for the probability that the observed

frequency X/n does not differ from 0.5 by more than 0.1;

iii) Determine the smallest value of n for which the probability that X/n does

not differ from 0.5 by more 0.1 is at least 0.95;

iv) If n = 50 and P(|(X/n) 0.5| < c) 0.9, determine the constant c. (Hint: In

(ii)(iv), utilize Tchebichevs inequality.)

5.2.3 Refer to Exercise 3.2.16 in Chapter 3 and suppose that 100 people are

chosen at random. Find the expected number of people with blood of each one

of the four types and the variance about these numbers.

5.2.4 If X is an r.v. distributed as P(), calculate the kth factorial moment

E[X(X 1) (X k + 1)].

5.2.5 Refer to Exercise 3.2.7 in Chapter 3 and find the expected number of

particles to reach the portion of space under consideration there during time

t and the variance about this number.

5.2.6 If X is an r.v. with a Hypergeometric distribution, use an approach

similar to the one used in the Binomial example in order to show that

(

mnr m + n r ) .

EX =

mr

, ( )

2 X =

m+n

(m + n) (m + n 1)

2

and p.

ii) By working as in the Binomial example, show that EX = rq/p, 2(X) =

rq/p2;

ii) Use (i) in order to show that EX = q/p and 2(X) = q/p2, if X has the

Geometric distribution.

5.2.8 Let f be the Gamma density with parameters = n, = 1. Then show

that

120 5 Moments of Random VariablesSome Moment and Probability Inequalities

n 1

x

f ( x)dx =

e .

x= 0 x!

Conclude that in this case, one may utilize the Incomplete Gamma tables (see,

for example, Tables of the Incomplete -Function, Cambridge University

Press, 1957, Karl Paerson, editor) in order to evaluate the d.f. of a Poisson

distribution at the points j = 1, 2, . . . .

5.2.9 Refer to Exercise 3.3.7 in Chapter 3 and suppose that each TV tube

costs $7 and that it sells for $11. Suppose further that the manufacturer sells an

item on money-back guarantee terms if the lifetime of the tube is less than c.

ii) Express his expected gain (or loss) in terms of c and ;

ii) For what value of c will he break even?

5.2.10 Refer to Exercise 4.1.12 in Chapter 4 and suppose that each bulb costs

30 cents and sells for 50 cents. Furthermore, suppose that a bulb is sold under

the following terms: The entire amount is refunded if its lifetime is <1,000 and

50% of the amount is refunded if its lifetime is <2,000. Compute the expected

gain (or loss) of the dealer.

5.2.11 If X is an r.v. having the Beta distribution with parameters and ,

then

ii) Show that

EX n =

( ) ( ),

+ +n

n = 1, 2, ;

( ) ( + + n)

ii) Use (i) in order to find EX and 2(X).

5.2.12 Let X be an r.v. distributed as Cauchy with parameters and 2. Then

show that E|X| = .

5.2.13 If the r.v. X is distributed as Lognormal with parameters and ,

compute EX, 2(X).

5.2.14 Suppose that the average monthly water consumption by the resi-

dents of a certain community follows the Lognormal distribution with = 104

cubic feet and = 103 cubic feet monthly. Compute the proportion of the

residents who consume more than 15 103 cubic feet monthly.

5.2.15 Let X be an r.v. with finite third moment and set = EX, 2 = 2(X).

Define the (dimensionless quantity, pure number) 1 by

3

X

1 = E .

1 is called the skewness of the distribution of the r.v. X and is a measure of

asymmetry of the distribution. If 1 > 0, the distribution is said to be skewed to

5.1 Moments of Random Exercises

Variables 121

the right and if 1 < 0, the distribution is said to be skewed to the left. Then show

that:

iii) If the p.d.f. of X is symmetric about , then 1 = 0;

iii) The Binomial distribution B(n, p) is skewed to the right for p < 1

2

and is

skewed to the left for p > 12 ;

iii) The Poisson distribution P() and the Negative Exponential distribution

are always skewed to the right.

5.2.16 Let X be an r.v. with EX4 < and define the (pure number) 2 by

4

X

2 = E

3, where = EX , 2 = 2 X . ( )

2 is called the kurtosis of the distribution of the r.v. X and is a measure of

peakedness of this distribution, where the N(0, 1) p.d.f. is a measure of

reference. If 2 > 0, the distribution is called leptokurtic and if 2 < 0, the

distribution is called platykurtic. Then show that:

ii) 2 < 0 if X is distributed as U(, );

ii) 2 > 0 if X has the Double Exponential distribution (see Exercise 3.3.13(iii)

in Chapter 3).

5.2.17 Let X be an r.v. taking on the values j with probability pj = P(X = j),

j = 0, 1, . . . . Set

()

G t = pj t j ,

j =0

1 t 1.

iii) Show that if |EX| < , then EX = d/dt G(t)|t = 1;

iii) Also show that if |E[X(X 1) (X k + 1)]| < , then

[ ( ) (

E X X 1 X k +1 = )] dk

dt k

Gt () t =1 ;

n j n j

p q , j 0, 0 < p < 1, q = 1 p

j

and

j

e , j 0, > 0;

j!

122 5 Moments of Random VariablesSome Moment and Probability Inequalities

iv) Utilize (ii) and (iii) in order to calculate the kth factorial moments of X

being B(n, p) and X being P(). Compare the results with those found in

Exercises 5.2.1 and 5.2.4, respectively.

If, in the preceding definitions, the p.d.f. f of the r. vector X is replaced by a

conditional p.d.f. f(xj , . . . , xj |xi , . . . , xi ), the resulting moments are called

1 n 1 m

Thus

x f x x ( )

(

)

2 2 1

E X 2 X1 = x1 = x 2

x f x x dx ,

2 2 1 2 ( )

[ ( )] f (x x )

2

x 2 E X 2 X1 = x1

( )

2 1

2 X 2 X1 = x1 = x 2

[ ( )] f (x x )dx .

2

x 2 E X 2 X1 = x1 2 1 2

For example, if (X1, X2) has the Bivariate Normal distribution, then f(x2 |x1)

is the p.d.f. of an N(b, 22 (1 2)) r.v., where

2

b = 2 +

1

(

x1 1 . )

Hence

(

E X 2 X1 = x1 = 2 + ) 2

1

(x1 1 . )

Similarly,

(

E X1 X 2 = x 2 = 1 + ) 1

2

(

x2 2 . )

Let X1, X2 be two r.v.s with joint p.d.f f(x1, x2). We just gave the definition

of E(X2|X1 = x1) for all x1 for which f(x2|x1) is defined; that is, for all x1 for which

fX (x1) > 0. Then E(X2|X1 = x1) is a function of x1. Replacing x1 by X1 and writing

1

E(X2|X1) instead of E(X2|X1 = x1), we then have that E(X2|X1 ) is itself an r.v.,

and a function of X1. Then we may talk about the E[E(X2|X1)]. In connection

with this, we have the following properties:

5.3 Conditional

5.1 Moments of Random Variables 123

(CE1) If E(X2) and E(X2|X1) exist, then E[E(X2|X1)] = E(X2) (that is,

the expectation of the conditional expectation of an r.v. is the

same as the (unconditional) expectation of the r.v. in question).

It suffices to establish the property for the continuous case only, for the

proof for the discrete case is quite analogous. We have

[( )] (

E E X 2 X1 = x 2 f x 2 x1 dx 2 fX x1 dx1

) 1

( )

( ) ( )

= x 2 f x 2 x1 fX x1 dx 2 dx1

1

( ) ( )

= x 2 f x1 , x 2 dx 2 dx1 = x 2 f x1 , x 2 dx1 dx 2

( )

= x 2 f x1 , x 2 dx1 dx 2 = x 2 fX x 2 dx 2 = E X 2 . ( ) ( )

REMARK 7 Note that here all interchanges of order of integration are legiti-

mate because of the absolute convergence of the integrals involved.

(CE2) Let X1, X2 be two r.v.s, g(X1) be a (measurable) function of X1

and let that E(X2) exists. Then for all x1 for which the

conditional expectations below exist, we have

[ ( ) ] ( )(

E X 2 g X1 X1 = x1 = g x1 E X 2 X1 = x1 )

or

[ ( ) ] ( )(

E X 2 g X1 X1 = g X1 E X 2 X1 . )

Again, restricting ourselves to the continuous case, we have

[ ( ) ] ( )( ) ( ) ( )

E X 2 g X1 X1 = x1 = x 2 g x1 f x 2 x1 dx 2 = g x1 x 2 f x 2 x1 dx 2

( ) (

= g x1 E X 2 X1 = x1 . )

In particular, by taking X2 = 1, we get

(CE2) For all x1 for which the conditional expectations below exist, we

have E[g(X1)|X1 = x1] = g(x1) (or E[g(X1)|X1] = g(X1)).

(CV) Provided the quantities which appear below exist, we have

[(

2 E X 2 X1 2 X 2)] ( )

and the inequality is strict, unless X2 is a function of X1 (on a set of

probability one).

Set ( ) ( )

= E X 2 , X1 = E X 2 X1 . ( )

124 5 Moments of Random VariablesSome Moment and Probability Inequalities

Then

( ) ( ) = E{[X (X )] + [ (X ) ]}

2 2

2 X2 = E X2 2 1 1

2 2

2 1 1 2 1 1

Next,

{[ ( )][ ( ) ]}

E X 2 X1 X1

= E[ X ( X )] E[ ( X )] E( X ) + E[ ( X )]

2 1

2

1 2 1

= E{E[ X ( X ) X ]} E[ ( X )] E[ E( X X )]

2 1 1

2

1 2 1

+ E[ ( X )] (by (CE1)),

1

[ ( )] [ ( )]

E 2 X1 E 2 X1 E X1 + E X1 [ ( )] [ ( )] (by (CE2)),

which is 0. Therefore

( ) [ ( )] [( ) ]

2 2

2 X 2 = E X 2 X1 + E X1 ,

and since

[ ( )]

2

E X 2 X1 0,

we have

( ) [( )

2 X 2 E X1 2 = 2 E X 2 X1 . ] [( )]

The inequality is strict unless

[ ( )]

2

E X 2 X1 = 0.

But

[ ( )] [ ( )] [ ( )]

2

E X 2 X1 = 2 X 2 X1 , since E X 2 X1 = = 0.

Remark 8, which follows.

Exercises

5.3.1 Establish properties (CEI) and (CE2) for the discrete case.

5.3.2 Let the r.v.s X, Y be jointly distributed with p.d.f. given by

5.4 Some Important Applications: Probability

5.1 Moments andof

Moment

Random

Inequalities

Variables 125

(

f x, y = ) 2

(

n n+1 )

if y = 1, . . . , x; x = 1, . . . , n, and 0 otherwise. Compute the following

(

quantities: E(X|Y = y), E(Y|X = x). Hint: Recall that nx =1 x = ( 2 ) , and

n n+1

x =1 x =

n 2 n ( )(

n + 1

6 )

2 n + 1)

.

5.3.3 Let X, Y be r.v.s with p.d.f. f given by f(x, y) = (x + y)I(0,1)(0,1)(x, y).

Calculate the following quantities: EX, 2(X), EY, 2(Y), E(X|Y = y),

2(X|Y = y).

5.3.4 Let X, Y be r.v.s with p.d.f. f given by f(x, y) = 2e(x+y)I(0,)(0,)(x, y).

Calculate the following quantities: EX, 2(X), EY, 2(Y), E(X|Y = y),

2(X|Y = y).

5.3.5 Let X be an r.v. with finite EX. Then for any r.v. Y, show that

E[E(X|Y)] = EX. (Assume the existence of all p.d.f.s needed.)

5.3.6 Consider the r.v.s X, Y and let h, g be (measurable) functions on ! into

itself such that E[h(X)g(Y)] and Eg(X) exist. Then show that

[( )( )

E h X gY X=x =h x E gY X=x. ] ()[( ) ]

THEOREM 1 Let X be a k-dimensional r. vector and g 0 be a real-valued (measurable)

function defined on !k, so that g(X) is an r.v., and let c > 0. Then

[ ( ) ] [ c ].

E g ( X)

Pg X c

[ ( )] ( )( )

E g X = g x1 , , xk f x1 , , xk dx1 dxk

( )( ) (

= g x1 , , xk f x1 , , xk dx1 dxk + c g x1 , , xk

A A

)

f ( x , , x )dx

1 k 1 dxk ,

126 5 Moments of Random VariablesSome Moment and Probability Inequalities

[ ( )] ( )( )

E g X g x1 , , xk f x1 , , xk dx1 dxk

A

c f ( x , , x )dx dx

1 k 1 k

A

Hence P[g(X) c] E[g(X)]/c. The proof is completely analogous if X is of the

discrete type; all one has to do is to replace integrals by summation signs.

1. Let X be an r.v. and take g(X) = |X |r, = E(X), r > 0. Then

r

EX

[ ]

P X c = P X c r

r

.

cr

2. In Markovs inequality replace r by 2 to obtain

( )=

2

EX 2 X

[ ]

2

P X c = P X c 2

2

= .

c2 c 2

c 2

[

P X k ] k

1

2

k

1

[

; equivalently, P X < k 1 2 . ]

REMARK 8 Let X be an r.v. with mean and variance 2 = 0. Then

Tchebichevs inequality gives: P[|X | c] = 0 for every c > 0. This result and

Theorem 2, Chapter 2, imply then that P(X = ) = 1 (see also Exercise 5.4.6).

LEMMA 1 Let X and Y be r.v.s such that

( )

E X = E Y = 0, ( ) ( )

2 X = 2 Y = 1.( )

Then

( )

E 2 XY 1 or, equivalently, 1 E XY 1, ( )

and

( )

E XY = 1 if any only if ( )

P Y = X = 1,

E( XY ) = 1 if any only if P(Y = X ) = 1.

PROOF We have

(

0 E X Y )

2

(

= E X 2 2 XY + Y 2 )

( )

= EX 2 2 E XY + EY 2 = 2 2 E XY ( )

5.4 Some Important Applications: Probability

5.1 Moments andof

Moment

Random

Inequalities

Variables 127

and

0 E X +Y( )

2

(

= E X 2 + 2 XY + Y 2 )

( )

= EX + 2 E XY + EY 2 = 2 + 2 E XY .

2

( )

Hence E(XY) 1 and 1 E(XY), so that 1 E(XY) 1. Now let P(Y = X)

= 1. Then E(XY) = EY2 = 1, and if P(Y = X) = 1, then E(XY) = EY2 = 1.

Conversely, let E(XY) = 1. Then

( ) ( ) [E(X Y )] = E(X Y )

2 2 2

2 X Y = E X Y

= EX 2 2 E( XY ) + EY = 1 2 + 1 = 0,

2

E(XY) = 1. Then 2(X + Y) = 2 + 2E(XY) = 2 2 = 0, so that

(

P X = Y = 1. )

THEOREM 2 (CauchySchwarz inequality) Let X and Y be two random variables with

means 1, 2 and (positive) variances 12 , 22 , respectively. Then

[( )(

E 2 X 1 Y 2 12 22 , )]

or, equivalently,

[(

1 2 E X 1 Y 2 1 2 , )( )]

and

[(

E X 1 Y 2 )( )] = 1 2

if and only if

P Y = 2 + 2 X 1 = 1

( )

1

and

[(

E X 1 Y 2 )( )] = 1 2

if and only if

P Y = 2 2 X 1 = 1.

1

( )

PROOF Set

X 1 Y 2

X1 = , Y1 = .

1 2

128 5 Moments of Random VariablesSome Moment and Probability Inequalities

(

E 2 X1Y1 1 )

if and only if

1 E X1Y1 1( )

becomes

[(

E 2 X 1 Y 2 )( )] 1

1

2 2

2

if and only if

[( )(

1 2 E X 1 Y 2 1 2 . )]

The second half of the conclusion follows similarly and will be left as an

exercise (see Exercise 5.4.6).

REMARK 9 A more familiar form of the CauchySchwarz inequality is

E2(XY) (EX2)(EY2). This is established as follows: Since the inequality is

trivially true if either one of EX2, EY2 is , suppose that they are both finite

and set Z = X Y, where is a real number. Then 0 EZ2 = (EX2)2

2[E(XY)] + EY2 for all , which happens if and only if E2(XY) (EX2)(EY2)

0 (by the discriminant test for quadratic equations), or E2(XY) (EX2)(EY2).

Exercises

5.4.1 Establish Theorem 1 for the discrete case.

5.4.2 Let g be a (measurable) function defined on ! into (0, ). Then, for any

r.v. X and any > 0,

Eg X( ).

[( ) ]

P g X

If furthermore g is even (that is, g(x) = g(x)) and nondecreasing for x 0, then

Eg X ( ).

(

P X ) g ( )

5.4.3 For an r.v. X with EX = and 2(X) = 2, both finite, use Tchebichevs

inequality in order to find a lower bound for the probability P(|X | < k).

Compare the lower bounds for k = 1, 2, 3 with the respective probabilities

when X N(, 2).

5.5 Covariance, Correlation

5.1 Coefficient

Moments of and

Random

Its Interpretation

Variables 129

. Use Tchebichevs inequality

in order to find a lower bound for the probability P(|(X/40) 1| 0.5),

and compare this bound with the exact value found from Table 3 in Appendix

III.

5.4.5 Refer to Remark 8 and show that if X is an r.v. with EX = (finite) such

that P(|X | c) = 0 for every c > 0, then P(X = ) = 1.

5.4.6 Prove the second conclusion of Theorem 2.

5.4.7 For any r.v. X, use the CauchySchwarz inequality in order to show

that E|X| E1/2X2.

In this section, we introduce the concepts of covariance and correlation coef-

ficient of two r.v.s and provide an interpretation for the latter. To this end,

for two r.v.s X and Y with means 1, 2, the (1, 1)-joint central mean, that is,

E[(X 1)(Y 2)], is called the covariance of X, Y and is denoted by

Cov(X,Y). If 1, 2 are the standard deviations of X and Y, which are assumed

to be positive, then the covariance of (X 1)/1, (Y 2)/2 is called the

correlation coefficient of X, Y and is denoted by (X, Y) or X,Y or 12 or just

if no confusion is possible; that is,

[( )(

X 1 Y 2 E X 1 Y 2

= E =

)] = Cov( X , Y )

1 2 1 2 1 2

=

( )

E XY 1 2

.

1 2

and = 1 if and only if

2

Y = 2 +

1

(

X 1 )

with probability 1, and = 1 if and only if

2

Y = 2

1

(

X 1 )

with probability 1. So = 1 means X and Y are linearly related. From this

stems the significance of as a measure of linear dependence between X and

Y. (See Fig. 5.1.) If = 0, we say that X and Y are uncorrelated, while if = 1,

we say that X and Y are completely correlated (positively if = 1, negatively if

= 1).

130 5 Moments of Random VariablesSome Moment and Probability Inequalities

Y

$2

#2 , $1 #1 $2

y ! #2 , $1 (x " # 1 )

#2

$2

y ! #2 " $1 (x " # 1 )

$2

#2 " $1 #1

X

0 #1

Figure 5.1

For 1 < < 1, 0, we say that X and Y are correlated (positively if > 0,

negatively if < 0). Positive values of may indicate that there is a tendency of

large values of Y to correspond to large values of X and small values of Y to

correspond to small values of X. Negative values of may indicate that small

values of Y correspond to large values of X and large values of Y to small

values of X. Values of close to zero may also indicate that these tendencies

are weak, while values of close to 1 may indicate that the tendencies are

strong.

The following elaboration sheds more light on the intuitive interpreta-

tion of the correlation coefficient (= (X, Y)) as a measure of co-linearity

of the r.v.s X and Y. To this end, for > 0, consider the line y = 2 + ( x 1 ) 2

1

in the xy-plane and let D be the distance of the (random) point (X, Y)

from the above line. Recalling that the distance of the point (x0, y0) from the

line ax + by + c = 0 is given by ax0 + by0 + c a 2 + b 2 . we have in the present

case:

1 12

D= X Y + 1 2 1 1+ ,

2 2 22

1

since here a = 1, b = and c = 1 2 1 . Thus,

2 2

2

12

D = X 1 Y + 1 2 1

2

1 + ,

2 2 22

and we wish to evaluate the expected squared distance of (X, Y) from the line

y = 2 + ( x 1 ) ; that is, ED2. Carrying out the calculations, we find

2

1

( 2

1 ) (

+ 22 D 2 = 22 X 2 + 12Y 2 2 1 2 XY + 2 2 1 2 2 1 X )

( ) ( )

2

2 1 1 2 2 1 Y + 1 2 2 1 . (1)

5.5 Covariance, Correlation

5.1 Coefficient

Moments of and

Random

Its Interpretation

Variables 131

EX 2 = 12 + 12 , EY 2 = 22 + 22 ( )

and E XY = 1 2 + 1 2 ,

we obtain

2 12 22

ED2 =

12 + 22

(1 ) ( > 0). (2)

2 12 22

ED2 =

12 + 22

(1+ ) ( < 0). (3)

2 + ( x 1 ) or y = 2 ( x 1 ), respectively (with probability 1). There-

2

1

2

1

fore, regardless of the value of , by the observation just made, relations (2)

and (3) are summarized by the expression

ED2 =

2 12 22

(

12 + 22

)

1 . (4)

relation (4) indicates the following: For > 0, the pairs (X, Y) tend to be

arranged along the line y = 2 + ( x 1 ) . These points get closer and closer

2

1

to this line as gets closer to 1, and lie on the line for = 1. For < 0, the

pairs (X, Y) tend to be arranged along the line y = 2 ( x 1 ) . These

2

1

points get closer and closer to this line as gets closer to 1, and lie on

this line for = 1. For = 0, the expected distance is constantly equal to

2 12 22 /( 12 + 22 ) from either one of the lines y = 2 + ( x 1 ) and

2

1

2

1

of the r.v.s X and Y.

The preceding discussion is based on the paper A Direct Development

of the Correlation Coefficient by Leo Katz, published in the American

Statistician, Vol. 29 (1975), page 170. His approach is somewhat different

and is outlined below. First, consider the r.v.s X1 and Y1 as defined in

the proof of Theorem 2; unlike the original r.v.s X and Y, the normalized

r.v.s X1 and Y1 are dimensionless. Through the transformations x1 = x 1

1

and y1 = y , we move from the xy-plane to the x1y1-plane. In this latter plane,

2

2

look at the point (X1, Y1) and seek the line Ax1 + By1 + C = 0

from which the expected squared distance of (X1, Y1) is minimum. That

is, determine the coefficients A, B and C, so that ED12 is minimum, where

132 5 Moments of Random VariablesSome Moment and Probability Inequalities

noticing that

we obtain

2 AB C2

ED12 = 1 + + . (5)

A2 + B2 A2 + B2

expression to be minimized is

2 AB

ED12 = 1 + . (6)

A2 + B2

(

A+ B )

2

( ) ( )

= A2 + B2 2 AB 0 A2 + B2 2 AB = A B , ( )

2

or equivalently,

2 AB

1 1. (7)

A2 + B2

If > 0, ED12 is minimized for A22AB+ B2

= 1 and the minimum is 1 .

If < 0, ED12 is minimized for A22AB+ B2

= 1 and the minimum is 1 + .

Finally, if = 0, the ED1 is constantly equal to 1; there is no minimizing

2

= 1 if and only if A =

B, and A22AB

+ B2

= 1 if and only if A = B. The corresponding lines are y1 = x1, the

main diagonal, and y1 = x1. Also observe that both minima of ED12 (for > 0

and < 0), and its constant value 1 (for = 0) are expressed by a single form,

namely, 1 ||.

To summarize: For > 0, the ED12 is minimized for the line y1 = x1; for

< 0, the ED12 is minimized for the line y1 = x1; for = 0, ED12 = 1, there is

no minimizing line. From this point on, the interpretation of as a measure of

co-linearity (of X1 and Y1) is argued as above, with the lines y = 2 + ( x 1 )

2

1

2

1

respectively.

5.1 Moments of Random Exercises

Variables 133

Exercises

5.5.1 Let X be an r.v. taking on the values 2, 1, 1, 2 each with probability

1 . Set Y = X2 and compute the following quantities: EX, 2(X), EY, 2(Y),

4

(X, Y).

5.5.2 Go through the details required in establishing relations (2), (3) and

(4).

5.5.3 Do likewise in establishing relation (5).

5.5.4 Refer to Exercise 5.3.2 (including the hint given there) and

iii) Calculate the covariance and the correlation coefficient of the r.v.s X and

Y;

iii) Referring to relation (4), calculate the expected squared distance of (X,

Y) from the appropriate line y = 2 + ( x 1 ) or y = 2 ( x 1 )

2

1

2

1

(which one?);

iii) What is the minimum expected squared distance of (X1, Y1) from

the appropriate line y = x or y = x (which one?) where X 1 = X and 1

1

Y1 = Y .

2

2

( ) .

2

n n n+1

Hint: Recall that

x = 2

3

x =1

5.5.5 Refer to Exercise 5.3.2 and calculate the covariance and the correla-

tion coefficient of the r.v.s X and Y.

5.5.6 Do the same in reference to Exercise 5.3.3.

5.5.7 Repeat the same in reference to Exercise 5.3.4.

constants and sgn x is 1 if x > 0 and is 1 if x < 0.

5.5.9 Let X and Y be r.v.s representing temperatures in two localities, A and

B, say, given in the Celsius scale, and let U and V be the respective tempera-

tures in the Fahrenheit scale. Then it is known that U and X are related as

follows: U = 59 X + 32, and likewise for V and Y. Fit this example in the model

of Exercise 5.5.8, and conclude that the correlation coefficients of X, Y and U,

V are identical, as one would expect.

5.5.10 Consider the jointly distributed r.v.s X, Y with finite second moments

and 2(X) > 0. Then show that the values and for which E[Y (X + )]2

is minimized are given by

134 5 Moments of Random VariablesSome Moment and Probability Inequalities

Y( ) X, Y .

= EY EX , =

(X )

( )

(The r.v. Y = X + is called the best linear predictor or Y, given X.)

5.5.11 If the r.v.s X1 and X2 have the Bivariate Normal distribution, show

that the parameter is, actually, the correlation coefficient of X1 and X2. (Hint:

Observe that the exponent in the joint p.d.f. of X1 and X2 may be written as

follows:

1 x 2 x 1 1 x 2 2 x 2 2

2

1 1

2 +

(

2 1 2 ) 1

1 2 2

(x ) (x )

2 2

1 b 2

=

1

2 2

+

2

2

, where b = 2 +

1

( )

x 1 1 .

1

2 2 1 2

5.5.12 If the r.v.s X1 and X2 have jointly the Bivariate Normal distribution

with parameters 1, 2, 12 , 22 and , calculate E(c1X1 + c2X2) and 2(c1X1 +

c2X2) in terms of the parameters involved, where c1 and c2 are real constants.

iii) If EX2 = EY2 < , then show that E(UV) = 0;

iii) If EX2, EY2 < and 2(X) = 2(Y), then U and V are uncorrelated.

moments, show that

m n m n

Cov X i , Yj = Cov X i , Yj .

i =1 j =1 i =1 j =1

( )

As a final application of the results of this chapter, we give a general proof of

Theorem 9, Chapter 2. To do this we remind the reader of the definition of the

concept of the indicator function of a set A.

5.6* Justification

5.1 Moments

of Relation

of Random

(2) in Chapter

Variables

2 135

denoted by IA, is a function on S defined as follows:

1 if s A

()

IA s =

0 if s A .

c

n

I I Aj = I A

n

j=1 j

(8)

j =1

n

I Aj = I A , (9)

n

j=1

j =1

j

and, in particular,

IA = 1 IA.

c (10)

Clearly,

( )

E IA = P A ( ) (11)

(1 X )(1 X ) (1 X ) = 1 H ( )

r

1 2 r 1 + H 2 + 1 H r , (12)

where Hj stands for the sum of the products Xi Xi , where the summation 1 j

extends over all subsets {i1, i2, . . . , ij} of the set {1, 2, . . . , r}, j = 1, . . . , r. Let

, be such that: 0 < , and + r. Then the following is true:

+

Xi

j

1

X i H J = ( )

H + , (13)

where J = {i1, . . . , i} is the typical member of all subsets of size of the set

{1, 2, . . . , r}, H(J) is the sum of the products Xj Xj , where the summation

1

extends over all subsets of size of the set {1, . . . , r} J, and J is meant to

The justification of (13) is as follows: In forming H+, we select ( + ) Xs

from the available r Xs in all possible ways, which is ( +r ). On the other hand,

for each choice of J, there are ( r ) ways of choosing Xs from the remaining

(r ) Xs. Since there are (r ) choices of J, we get (r )( r ) groups (products)

of ( + ) Xs out of r Xs. The number of different groups of ( + ) Xs out

136 5 Moments of Random VariablesSome Moment and Probability Inequalities

number of distinct ones is given by

r r r! r ! ( )

(

! r ! ! r !

=

) ( )

r r!

+

+ ! (

r ! )( )

=

( + )! = + .

! !

Now clearly,

Bm = Ai Ai Aic Aic ,

1 m m +1 M

Jm

where the summation extends over all choices of subsets Jm = {i1, . . . , im} of the

set {1, 2, . . . , M} and Bm is the one used in Theorem 9, Chapter 2. Hence

IB = IA

m i1 Aim Aicm + 1 AicM

Jm

= IA IA 1 IA

Jm

i1 im ( im + 1 ) (1 I ) (by (8), (9), (10))

AiM

( )

= I A I A 1 H1 J m + H 2 J m + 1 ( ) ( ) ( )

H M m J m

M m

J m

i1

im

(by (12)).

Since

IA

Jm

i1 im

m + k

I A H k Jm = ( )

m

H m+ k (by (13)),

we have

m + 1 m + 2 M

( )

M m

I B = Hm H m+1 + H m+ 2 + 1 HM .

m

m m m

Taking expectations of both sides, we get (from (11) and the definition of Sr in

Theorem 9, Chapter 2)

m + 1 m + 2 M

( ) ( )

M m

P Bm = S m S m+1 + S m+ 2 + 1 SM ,

m m m

as was to be proved.

5.6* Justification

5.1 Moments

of Relation

of Random

(2) in Chapter

Variables

2 137

(For the proof just completed, also see pp. 8085 in E. Parzens book

Modern Probability Theory and Its Applications published by Wiley, 1960.)

REMARK 10 In measure theory the quantity IA is sometimes called the char-

acteristic function of the set A and is usually denoted by A. In probability

theory the term characteristic function is reserved for a different concept and

will be a major topic of the next chapter.

138 6 Characteristic Functions, Moment Generating Functions and Related Theorems

Chapter 6

Characteristic Functions,

Moment Generating Functions

and Related Theorems

6.1 Preliminaries

The main subject matter of this chapter is the introduction of the concept of

the characteristic function of an r.v. and the discussion of its main properties.

The characteristic function is a powerful mathematical tool, which is used

profitably for probabilistic purposes, such as producing the moments of an r.v.,

recovering its distribution, establishing limit theorems, etc. To this end, recall

that for z !, eiz = cos z + i sin z, i = 1 , and in what follows, i may

be treated formally as a real number, subject to its usual properties: i2 = 1,

i3 = i, i 4 = 1, i 5 = i, etc.

The sequence of lemmas below will be used to justify the theorems which

follow, as well as in other cases in subsequent chapters. A brief justification for

some of them is also presented, and relevant references are given at the end of

this section.

( ) ( )

g1 x j g 2 x j , j = 1, 2, ,

PROOF If the summations are finite, the result is immediate; if not, it follows

by taking the limits of partial sums, which satisfy the inequality. "

b

LEMMA A Let g1, g2 : ! [0, ) be such that g1(x) g2(x), x !, and that a g1 (x)dx

exists for every a, b, ! with a < b, and that g 2 (x)dx < . Then g1(x)dx

< .

138

6.5 The Moment Generating

6.1 Preliminaries

Function 139

LEMMA B Let g : {x1, x2, . . .} ! and xj|g(xj)| < . Then xj g(xj) also converges.

PROOF The result is immediate for finite sums, and it follows by taking the

limits of partial sums, which satisfy the inequality. "

b

LEMMA B Let g : ! ! be such that a g(x)dx exists for every a, b, ! with a < b, and that

|g( x ) |dx < . Then g(x)dx also converges.

PROOF Same as above replacing sums by integrals. "

The following lemma provides conditions under which the operations

of taking limits and expectations can be interchanged. In more advanced

probability courses this result is known as the Dominated Convergence

Theorem.

LEMMA C Let {Xn}, n = 1, 2, . . . , be a sequence of r.v.s, and let Y, X be r.v.s such

that |Xn(s)| Y(s), s S, n = 1, 2, . . . and Xn(s) X(s) (on a set of ss of

probability 1) and E(Y) < . Then E(X) exists and E(Xn) n

E(X), or

equivalently,

( ) (

lim E X n = E lim X n .

n n

)

REMARK 1 The index n can be replaced by a continuous variable.

The next lemma gives conditions under which the operations of differen-

tiation and taking expectations commute.

LEMMA D For each t T (where T is ! or an appropriate subset of it, such as the interval

[a, b]), let X(; t) be an r.v. such that (t)X(s; t) exists for each s S and

t T. Furthermore, suppose there exists an r.v. Y with E(Y) < and such

that

t

( )

X s; t Y s ,() s S , t T .

Then

d

dt [ ( )]

( )

E X ; t = E X ; t , for all t T .

t

The proofs of the above lemmas can be found in any book on real vari-

ables theory, although the last two will be stated in terms of weighting func-

tions rather than expectations; for example, see Advanced Calculus, Theorem

2, p. 285, Theorem 7, p. 292, by D. V. Widder, Prentice-Hall, 1947; Real

Analysis, Theorem 7.1, p. 146, by E. J. McShane and T. A. Botts, Van

Nostrand, 1959; The Theory of Lebesgue Measure and Integration, pp. 6667,

by S. Hartman and J. Mikusinski, Pergamon Press, 1961. Also Mathematical

Methods of Statistics, pp. 4546 and pp. 6668, by H. Cramr, Princeton

University Press, 1961.

140 6 Characteristic Functions, Moment Generating Functions and Related Theorems

Let X be an r.v. with p.d.f. f. Then the characteristic function of X (ch.f. of X),

denoted by X (or just when no confusion is possible) is a function defined on

!, taking complex values, in general, and defined as follows:

() [ ( ) ( ) ( ) ( )]

e itx f x = cos tx f x + i sin tx f x

()

X t = E e itX [ ]

= x x

() [ ( ) ( ) ( ) ( )]

e f x dx =

cos tx f x + i sin tx f x dx

itx

[ ( ) ( )] [ ( ) ( )]

cos tx f x + i sin tx f x

= x x

( )() ( )()

cos tx f x dx + i sin tx f x dx.

By Lemmas A, A, B, B, X(t) exists for all t !. The ch.f. X is also called the

Fourier transform of f.

The following theorem summarizes the basic properties of a ch.f.

THEOREM 1 (Some properties of ch.fs)

i) X(0) = 1.

ii)|X(t)| 1.

iii)X is continuous, and, in fact, uniformly continuous.

iv) X+d(t) = eitdX(t), where d is a constant.

v) cX(t) = X(ct), where c is a constant.

vi) cX+d(t) = eitdX(ct).

dn

vii)

dt n

X t () = inE(Xn), n = 1, 2, . . . , if E|Xn| < .

t =0

PROOF

i) X(t) = EeitX. Thus X(0) = Eei0X = E(1) = 1.

ii) |X(t)| = |EeitX| E|eitX| = E(1) = 1, because |eitX| = 1. (For the proof of the

inequality, see Exercise 6.2.1.)

( )

iii) X t + h X t = E e

() ( )

i t+ h X

e itX

[ ( )] ( )

= E e itX e ihX 1 E e itX e ihX 1

= E e ihX 1.

Then

6.2 Definitions and Basic TheoremsThe

6.5 The Moment

One-Dimensional

Generating Function

Case 141

h0

( ) ()

lim X t + h X t lim E e ihX 1 = E lim e ihX 1 = 0,

h0 [ h0 ]

provided we can interchange the order of lim and E, which here can be

done by Lemma C. We observe that uniformity holds since the last ex-

pression on the right is independent of t.

iv) X+d(t) = Eeit(X+d) = E(eitXeitd) = eitd EeitX = eitd X(t).

v) cX(t) = Eeit(cX) = Eei(ct)X = X(ct).

vi) Follows trivially from (iv) and (v).

n

vii)

dn

dt n X

dn

()

t = n Ee itX = E n e itX = E i n X n e itX ,

dt t

( )

provided we can interchange the order of differentiation and E. This can

be done here, by Lemma D (applied successively n times to X and its

n 1 first derivatives), since E|X n| < implies E|X k| < , k = 1, . . . , n

(see Exercise 6.2.2). Thus

dn

dt n

X t () ( )

= inE X n . "

t =0

REMARK

n

2 From part (vii) of the theorem we have that E(Xn) =

(i) n d

dt n

X(t)|t = 0, so that the ch.f. produces the nth moment of the r.v.

REMARK 3 If X is an r.v. whose values are of the form x = a + kh, where a,

h are constants, h > 0, and k runs through the integral values 0, 1, . . . , n or 0,

1, . . . , or 0, 1, . . . , n or 0, 1, . . . , then the distribution of X is called a lattice

distribution. For example, if X is distributed as B(n, p), then its values are of

the form x = a + kh with a = 0, h = 1, and k = 0, 1, . . . , n. If X is distributed as

P(), or it has the Negative Binomial distribution, then again its values are of

the same form with a = 0, h = 1, and k = 0, 1, . . . . If now is the ch.f. of X, it

can be shown that the distribution of X is a lattice distribution if and only if

|(t)| = 1 for some t 0. It can be readily seen that this is indeed the case in the

cases mentioned above (for example, (t) = 1 for t = 2). It can also be shown

that the distribution of X is a lattice distribution, if and only if the ch.f. is

periodic with period 2 (that is, (t + 2) = (t), t !).

In the following result, the ch.f. serves the purpose of recovering the

distribution of an r.v. by way of its ch.f.

THEOREM 2 (Inversion formula) Let X be an r.v. with p.d.f. f and ch.f. . Then if X is of the

discrete type, taking on the (distinct) values xj, j 1, one has

( ) 1

()

T itx j

i) f x j = lim

T 2T

T

e t dt , j 1.

142 6 Characteristic Functions, Moment Generating Functions and Related Theorems

1 e ith itx

() 1

()

T

ii) f x = lim lim

h0 T 2 T ith

e t dt

and, in particular, if | (t )|dt < , then (f is bounded and continuous and)

1 itx

ii) f(x) = e (t)dt.

2

PROOF (outline) i) The ch.f. is continuous, by Theorem 1(iii), and since

T itx

so is eitxj, it follows that the integral T e (t)dt exists for every T(> 0). We j

have then

itx

1

() 1

( )

T T

e itx f x k dt

itx j

e t dt =

e

j k

2T T 2T T

k

(

) f x dt

1

( )

T

e

it x x

2T T k

= k

k j

1 T it( x x )

= f xk

k

( )

2T T

e dt k j

(the interchange of the integral and summations is valid here). That is,

( ) dt .

1

() ( ) 21T

T T

t dt = f x k

itx j it x k x j

2T T e k

T

e (1)

But

( ) dt =

[cos t(x ) ( )]

T it x k x j T

T

e T

k x j + i sin t x k x j dt

( ) ( )

T T

= cos t x k x j dt + i sin t x k x j dt

T T

T

k j

T

( ) dt = cos t x

( x )dt .

T it x k x j T

T

e T

k j (2)

( ) 1

( )

T T

T

cos t x k x j dt =

xk x j T

d sin t x k x j

=

( )

sin T x k x j sin T x k x j [ ( )]

xk x j

=

(

2 sin T x k x j ).

xk x j

6.2 Definitions and Basic TheoremsThe

6.5 The Moment

One-Dimensional

Generating Function

Case 143

Therefore,

1, if xk = x j

( ) dt = sin T x x

1

( k j)

T it x k x j

2T

T

e

, if xk x j.

(3)

(

T x k x j )

sinT ( x x )

But k

x x

j

x 1 x , a constant independent of T, and therefore, for

k j

sin T ( x x )

k j

k j

T k j

1 T (

it x k x j ) dt = 1, if xk = x j

lim

T 2T e

xk x j.

(4)

0, if

T

1

() ( ) 21T ( )

T T

t dt = lim f xk

it xk x j

T e

itx j

lim e dt

T 2T T

k

T

( ) 1 ( )

T

= f xk lim

it xk x j

k

T 2T T e dt

( ) dt = f x ,

( ) 1

( j)

T

= f x j + lim

it x k x j

k j

T 2T T e

as was to be seen

ii) (ii) Strictly speaking, (ii) follows from (ii). We are going to omit (ii)

entirely and attempt to give a rough justification of (ii). The assumption that

itx

| (t )|dt < implies that e (t)dt exists, and is taken as follows for every

arbitrary but fixed x !:

() ()

T

e itx t dt = lim e itx t dt .

(0 < )T T

(5)

But

()

e itx t dt = e itx e ity f y dydt ()

T T

T T

() ()

it ( y x ) it ( y x )

f y dy dt = f y e dt dy,

T T

= e (6)

T T

where the interchange of the order of integration is legitimate here. Since the

integral (with respect to y) is zero over a single point, we may assume in the

sequel that y x. Then

T (

it y x ) 2 sin T y x ( ),

T e dt =

yx

(7)

144 6 Characteristic Functions, Moment Generating Functions and Related Theorems

sin T y x ( ) dy.

() ()

e t dt = 2 lim f y (8)

itx

T yx

Setting T(y x) = z, expression (8) becomes

z sin z

()

e itx t dt = 2 lim f x +

T T z

dz

()

= 2 f x = 2f x , ()

by taking the

limit under the integral sign, and by using continuity of f and the

fact that sinz z dz = . Solving for f(x), we have

() 1

()

f x =

2

e itx t dt ,

as asserted. "

EXAMPLE 1 Let X be B(n, p). In the next section, it will be seen that X(t) = (peit + q)n. Let

us apply (i) to this expression. First of all, we have

1

() 1

( pe )

T T n

e itx t dt = + q e itx dt

it

2T T 2T T

n

r ( pe )q

n

1 T r

n r

= it

e itx dt

2T T

r = 0

1 n

n T ( )

r p q n r i r x t

= r

e dt

2T r= 0

T

1 n

n 1 ( )

( )

T

r p q

i r x t

n r

= r

e i r x dt

2T r= 0 i( r x ) T

r x

1 n x n x T

+ p q T dt

2T x

i ( r x )T i ( r x )T

n

n e e 1 n x n x

= p r qn r + p q 2T

r= 0 r

r x

2Ti r x 2T x ( )

n

n

= p r qn r

sin r x T n x n x

+ p q .

( )

r= 0 r

r x

rx T x ( )

Taking the limit as T , we get the desired result, namely

n

()

f x = p x q n x .

x

6.5 The Moment GeneratingExercises

Function 145

(One could also use (i) for calculating f(x), since is, clearly, periodic with

period 2.)

EXAMPLE 2 For an example of the continuous type, let X be N(0, 1). In the next section, we

will see that X(t) = et /2. Since |(t)| = et /2, we know that | (t )|dt < , so that

2 2

1 1

( ) ()

2

f x = e itx t dt = e itx e t 2

dt

2 2

1 ( 1 2 )( t + 2 itx ) 1 ( 1 2 ) t + 2 t ( ix ) + ( ix ) ( 1 2 )( ix )

2 2 2

2

2 2

= e dt = e e dt

( 1 2 )x 2

( 1 2 )x 2

e 1 ( 1 2 )( t + ix ) e 2

1 ( 1 2 )u2

=

2

2

e dt =

2

2

e du

( )

1 2 x2

e 1 2

= 1 = e x 2

,

2 2

as was to be shown.

characteristic function and the p.d.f. of a random variable.

PROOF The p.d.f. of an r.v. determines its ch.f. through the definition of the

ch.f. The converse, which is the involved part of the theorem, follows from

Theorem 2. "

Exercises

6.2.1 Show that for any r.v. X and every t !, one has |EeitX| E|eitX|(= 1).

(Hint: If z = a + ib, a, b !, recall that z = a 2 + b 2 . Also use Exercise 5.4.7

in Chapter 5 in order to conclude that (EY)2 EY2 for any r.v. Y.)

6.2.2 Write out detailed proofs for parts (iii) and (vii) of Theorem 1 and

justify the use of Lemmas C, D.

6.2.3 For any r.v. X with ch.f. X, show that X(t) = X(t), t !, where the bar

over X denotes conjugate, that is, if z = a + ib, a, b !, then z = a ib.

6.2.4 Show that the ch.f. X of an r.v. X is real if and only if the p.d.f. fX of X

is symmetric about 0 (that is, fX(x) = fX(x), x !). (Hint: If X is real, then the

conclusion is reached by means of the previous exercise and Theorem 2. If fX

is symmetric, show that fX(x) = fX(x), x !.)

146 6 Characteristic Functions, Moment Generating Functions and Related Theorems

6.2.5 Let X be an r.v. with p.d.f. f and ch.f. given by: (t) = 1 |t| if |t| 1

and (t) = 0 if |t| > 1. Use the appropriate inversion formula to find f.

6.2.6 Consider the r.v. X with ch.f. (t) = e|t|, t !, and utilize Theorem 2(ii)

in order to determine the p.d.f. of X.

In this section, the ch.f.s of some distributions commonly occurring will be

derived, both for illustrative purposes and for later use.

1. Let X be B(n, p). Then X(t) = (peit + q)n. In fact,

n n

( ) ( )

n n

()

x n

X t = e itx p x q n x = pe it q n x = pe it + q ,

x= 0 x x = 0 x

Hence

d

() ( )

n 1

X t = n pe it + q ipe it = inp,

dt t=0 t=0

d2

() (

d it

)

n 1

X t = inp pe + q e it

dt 2 t=0

dt t= 0

( )( ) ( )

n 2 n 1

= inp n 1 pe it + q ipe it e it + pe it + q ie it

t= 0

[( ) ]

= i 2 np n 1 p + 1 = np n 1 p + 1 = E X 2 ,[( ) ] ( )

so that

( ) [( ) ]

E X 2 = np n 1 p + 1 and 2 X = E X 2 EX ( ) ( ) ( ) 2

= n 2 p 2 np 2 + np n 2 p 2 = np 1 p = npq; ( )

that is, 2(X ) = npq.

2. Let X be P(). Then X(t) = ee . In fact,

it

( )

x

e

it

x

()

X t = e itx e =e

it it

= e e e = e e

.

x= 0 x! x= 0 x!

6.3 The Characteristic 6.5

Functions

The Moment

of Some

Generating

Random Variables

Function 147

Hence

d

()

it

X t = e e

ie it = i ,

dt t=0 t= 0

d2

dt 2

X t()

t=0

=

d

dt

(

ie e e

it

+it

) t=0

d e it

= ie e +it

dt t= 0

= ie

e eit +it

(

e it i + i ) t= 0

= ie ( )

e i + i

= i ( + 1) = ( + 1) = E ( X ),

2 2

so that

( )

2 X = E X 2 EX ( ) ( ) 2

(

= + 1 2 = ; )

that is, 2(X ) = .

1. Let X be N(, 2). Then X(t) = eit( t /2), and, in particular, if X is

22

N(0, 1), then X(t) = et /2. If X is N(, 2), then (X )/, is N(0, 1).

2

Thus

() () ( )

( X ) t = (1 ) X ( ) t = e it X t , and X t = e it ( X ) t . ( ) ()

So it suffices to find the ch.f. of an N(0, 1) r.v. Y, say. Now

( ) 2 dy

() 1 1

y 2 2 ity

e e

2

Y t = ity

e y 2

dy =

2 2

1 (

y it )

2

2

e

2 2

= e t 2

dy = e t 2

.

2

2

2t 2

()

X t = exp it

2

.

148 6 Characteristic Functions, Moment Generating Functions and Related Theorems

Hence

2t 2

d

dt

()

X t = exp it

2 i t

2

( ) = i , so that E X = . ( )

t =0 t =0

2t 2 2t 2

d2

() ( )

2

X t = exp it i t

2

2 exp it

dt 2 t =0

2 2

t =0

= i2 2 2 = i2 2 + 2 . ( )

Then E(X 2) = 2 + 2 and 2(X) = 2 + 2 2 = 2.

2. Let X be Gamma distributed with parameters and . Then X(t) =

(1 it). In fact,

() 1 1 ( )

x 1 it

( ) 0 ( ) 0

X t = e itx x 1 e x dx = x 1 e dx.

Setting x(1 it) = y, we get

x=

y

1 it

, dx =

dy

1 it

, [

y 0, . )

Hence the above expression becomes

1 1 dy

( ) 0 (1 it ) 1

1 y

y e

1 it

( ) 1

( )

( ) 0

= 1 it y 1e y dy = 1 it .

Therefore

i

d

X t () = = i ,

(1 it )

+1

dt

t =0 t =0

d2 (

+1 2 )

X =i 2

(

= i 2 + 1 2 , )

(1 it )

+2

dt 2

t =0 t =0

2 2 2

For = r/2, = 2, we get the corresponding quantities for 2r, and for

= 1, = 1/, we get the corresponding quantities for the Negative Exponential

distribution. So

6.5 The Moment GeneratingExercises

Function 149

1

it

() ( ) ()

r 2

X t = 1 2it , X t = 1 = ,

it

respectively.

3. Let X be Cauchy distributed with = 0 and = 1. Then X(t) = e|t|. In

fact,

1 cos tx ( )

() 1 1

X t = e itx dx = dx

1+ x 2

1 + x 2

i sin tx

+ dx =

( )

2 cos tx

dx

( )

1+ x 2

0 1 + x2

because

( ) dx = 0,

sin tx

1 + x2

since sin(tx) is an odd function, and cos(tx) is an even function. Further, it can

be shown by complex variables theory that

cos tx ( ) dx = e t

0 1 + x2 2

.

Hence

()

X t = e

t

.

Now

d

dt

X t = e

d t

dt

()

does not exist for t = 0. This is consistent with the fact of nonexistence of E(X ),

as has been seen in Chapter 5.

Exercises

6.3.1 Let X be an r.v. with p.d.f. f given in Exercise 3.2.13 of Chapter 3.

Derive its ch.f. , and calculate EX, E[X(X 1)], 2(X), provided they are

finite.

6.3.2 Let X be an r.v. with p.d.f. f given in Exercise 3.2.14 of Chapter 3.

Derive its ch.f. , and calculate EX, E[X(X 1)], 2(X), provided they are

finite.

6.3.3 Let X be an r.v. with p.d.f. f given by f(x) = e(x) I(,)(x). Find its ch.f.

, and calculate EX, 2(X), provided they are finite.

150 6 Characteristic Functions, Moment Generating Functions and Related Theorems

and p.

i) Show that its ch.f., , is given by

pr

()

t = ;

( )

r

1 qe it

iii) Find the quantities mentioned in (i) and (ii) for the Geometric

distribution.

6.3.5 Let X be an r.v. distributed as U(, ).

ii) Show that its ch.f., , is given by

e it e it

()

t = ;

(

it )

ii) By differentiating , show that EX = 2+ and 2 ( X ) = ( )2

12

.

6.3.6 Consider the r.v. X with p.d.f. f given in Exercise 3.3.14(ii) of Chapter

3, and by using the ch.f. of X, calculate EX n, n = 1, 2, . . . , provided they are

finite.

In this section, versions of Theorems 1, 2 and 3 are presented for the case that

the r.v. X is replaced by a k-dimensional r. vector X. Their interpretation,

usefulness and usage is analogous to the ones given in the one-dimensional

case. To this end, let now X = (X1, . . . , Xk) be a random vector. Then the ch.f.

of the r. vector X, or the joint ch.f. of the r.v.s X1, . . . , Xk, denoted by X or

X1, . . . , Xk, is defined as follows:

X , ,

1 Xk (t , , t ) = E[e

1 k

it 1 X 1 +it2 X 2 + + itk X k

], t j ! ,

of Lemmas A, A and B, B. The joint ch.f. X1, . . . , Xk satisfies properties

analogous to properties (i)(vii). That is, one has

THEOREM 1 (Some properties of ch.f.s)

i) X1, . . . , Xk(0, . . . , 0) = 1.

ii) |X1, . . . , Xk(t1, . . . , tk)| 1.

iii) X1, . . . , Xk is uniformly continuous.

6.4 Definitions and Basic TheoremsThe

6.5 The Moment

Multidimensional

Generating Function

Case 151

v) c1X1, . . . , ckXk(t1, . . . , tk) = X1, . . . , Xk (c1t1, . . . , cktk).

it1d1 + + itkdk

vi) c1X1+ d1, . . . , ckXk + dk(t1, . . . , tk) = e X1, . . . , Xk(c1t1, . . . , cktk).

vii) If the absolute (n1, . . . , nk)-joint moment, as well as all lower order joint

moments of X1, . . . , Xk are finite, then

n + +n

(t , , t ) = i ( )

1 k k

j=1 n j

X , , Xk 1 k E X 1n X kn ,

1 k

t tn t tn

1 k 1

1 k t 1 = =tk = 0

and, in particular,

n

t nj

X , , 1 Xk (t , , t )

1 k ( )

= i n E X nj , j = 1, 2, , k.

t 1 = =tk = 0

viii) If in the X1, . . . , Xk(t1, . . . , tk) we set tj1 = = tjn = 0, then the resulting

expression is the joint ch.f. of the r.v.s Xi1, . . . , Xim, where the js and the

is are different and m + n = k.

Multidimensional versions of Theorem 2 and Theorem 3 also hold true.

We give their formulations below.

THEOREM 2 (Inversion formula) Let X = (X1, . . . , Xk) be an r. vector with p.d.f. f and ch.f.

. Then

k

1

( )

T T

ii) f X , ,

1 Xk x1 , , xk = lim

T 2T T

e it x it x

T

1 1 k k

X , ,

1 Xk (t , , t )dt

1 k 1 dt k ,

k

1 k 1 e it j h

(x , , x )

T T

ii) f X , ,

1 Xk 1 k = lim lim

h 0 T 2 T

T

j =1 it j h

1 k 1 dt k ,

if X is of the continuous type, with the analog of (ii) holding if the integral

of |X1, . . . , Xk(t1, . . . , tk)| is finite.

THEOREM 3 (Uniqueness Theorem) There is a one-to-one correspondence between the

ch.f. and the p.d.f. of an r. vector.

PROOFS The justification of Theorem 1 is entirely analogous to that given

for Theorem 1, and so is the proof of Theorem 2. As for Theorem 3, the fact

that the p.d.f. of X determines its ch.f. follows from the definition of the ch.f.

That the ch.f. of X determines its p.d.f. follows from Theorem 2. "

152 6 Characteristic Functions, Moment Generating Functions and Related Theorems

Let X = (X1, . . . , Xk) be Multinomially distributed; that is,

(

P X 1 = x1 , , X k = xk = ) n!

x1! xk !

p1x pkx . 1 k

Then

(t , , t ) = ( p e ).

n

it 1

X , ,

1 Xk 1 k 1 + + pk e it k

In fact,

X , ,

1 Xk (t , , t ) =

1 k

x 1 , , xk

e it x + +it x

1 1 k k

n!

x1! xk !

p1x pkx 1 k

n!

( ) ( )

x1 xk

= x1! xk !

p1e it 1

pk e it k

x 1 , , xk

( ).

n

= p1e it + + pk e it

1 k

Hence

k

t1 t k

X , ,1 Xk (t , , t )

1 k

t 1 = = tk = 0

( ) ( )

= n n 1 n k + 1 i p1 pk p1e it1 + k

(

) ( ) ( )

n k

+ pk e itk = i k n n 1 n k + 1 p1 p2 pk .

t 1 = = tk = 0

Hence

( ) (

E X1 X k = n n 1 n k + 1 p1 p2 pk .) ( )

Finally, the ch.f. of a (measurable) function g(X) of the r. vector X =

(X1, . . . , Xk) is defined by:

e itg ( x ) f x , x = x ,

()

1 , xk ( )

()

g(X) t = E e

itg ( X )

= x

( )

itg ( x , , x )

e f x1 , , xk dx1 , , dxk .

1 k

Exercise

6.4.1 (CramrWold) Consider the r.v.s Xj, j = 1, . . . , k and for cj !,

j = 1, . . . , k, set

6.5 The Moment Generating Function 153

k

Yc = c j X j .

j =1

Then

ii) Show that Yc(t) = X1, . . . , Xk(c1t, . . . , ckt), t !, and X1, . . . , Xk(c1, . . . , ck)

= Yc(1);

ii) Conclude that the distribution of the Xs determines the distribution of Yc

for every cj !, j = 1, . . . , k. Conversely, the distribution of the Xs is

determined by the distribution of Yc for every cj !, j = 1, . . . , k.

Generating Function of a Random Variable

The ch.f. of an r.v. or an r. vector is a function defined on the entire real line

and taking values in the complex plane. Those readers who are not well versed

in matters related to complex-valued functions may feel uncomfortable in

dealing with ch.f.s. There is a partial remedy to this potential problem, and

that is to replace a ch.f. by an entity which is called moment generating

function. However, there is a price to be paid for this: namely, a moment

generating function may exist (in the sense of being finite) only for t = 0. There

are cases where it exists for ts lying in a proper subset of ! (containing 0), and

yet other cases, where the moment generating function exists for all real t. All

three cases will be illustrated by examples below.

First, consider the case of an r.v. X. Then the moment generating function

(m.g.f.) MX (or just M when no confusion is possible) of a random variable X,

which is also called the Laplace transform of f, is defined by MX(t) = E(etX),

t !, if this expectation exists. For t = 0, MX(0) always exists and equals

1. However, it may fail to exist for t 0. If MX(t) exists, then formally

X(t) = MX(it) and therefore the m.g.f. satisfies most of the properties analo-

gous to properties (i)(vii) cited above in connection with the ch.f., under

suitable conditions. In particular, property (vii) in Theorem 1 yields

M X (t ) t =0 = E ( X n ) , provided Lemma D applies. In fact,

n

d

n

dt

dn

dn

dt n

MX t () =

dn

dt n

(

Ee tX ) = E n e tX

dt t =0

t =0 t =0

(

= E X n e tX ) t =0

( )

= E Xn .

This is the property from which the m.g.f. derives its name.

154 6 Characteristic Functions, Moment Generating Functions and Related Theorems

to see how conditions are imposed on t in order for the m.g.f. to be finite. It so

happens that part (vii) of Theorem 1, as it would be formulated for an m.g.f.,

is applicable in all these examples, although no justification will be supplied.

1. If X B(n, p), then MX(t) = (pet + q)n, t !. Indeed,

n n

( ) ( )

n n

()

x n

M X t = e tx p x q n x = pe t q n x = pe t + q ,

x= 0 x x = 0 x

Then

d

() d

( ) ( ) ( )

n n 1

MX t = pe t + q = n pe t + q pe t = np = E X ,

dt t =0

dt t =0 t =0

and

d2

() d

( )

n 1

MX t = np pe t + q et

dt 2 t =0

dt t =0

( )( ) ( )

n 2 n 1

= np n 1 pe t + q pe t e t + pe t + q et

t =0

( )

= n n 1 p 2 + np = n 2 p 2 np 2 + np = E X 2 , ( )

so that 2(X) = n2p2 np2 + np n2p2 = np(1 p) = npq.

2. If X P(), then MX(t) = ee , t !. In fact,

t

( )

x

e

t

x

()

M X t = e tx e =e

t t

= e e e = e e .

x= 0 x! x= 0 x!

Then

d

() d e

( )

t t

MX t = e = e t e e ==E X ,

dt t =0

dt t =0

t =0

and

d2

dt 2

()

MX t

t =0

=

d

dt

( t

e t e e ) t =0

( t

= e t e e + e t e e e t

t

) t =0

( ) ( )

= 1 + = E X 2 , so that 2 X = + 2 2 = . ( )

6.5 The Moment Generating Function 155

2t 2

3. If X N(, 2), then M X (t ) = e

t +

2

, t !, and, in particular, if X N(0,

1), then MX(t) = et /2, t !. By the property for m.g.f. analogous to property (vi)

2

in Theorem 1,

t t

()

M X t = e t M X , so that M X = e t M X t

()

t t t 2

t 2

2

+

MX = e et =e 2

.

2t 2

Replacing t by t, we get, finally, M X (t ) = e

t +

2

. Then

2t 2 2t 2

d

() d t +

( ) ( )

t +

MX t = e 2

= + 2t e 2

==E X ,

dt t= 0

dt

t=0 t= 0

and

d

2t 2

d2

() ( )

t +

MX t = + 2t e 2

dt 2 dt

t=0 t= 0

t 2 2

t 2 2

( )

t + 2 t +

= 2 e 2

+ + 2t e 2

= 2 + 2

t=0

( )

= E X 2 , so that 2 X = 2 + 2 2 = 2 . ( )

4. If X is distributed as Gamma with parameters and , then MX(t) =

(1 t), t < 1/. Indeed,

() 1 1 ( )

x 1 t

( ) ( )

MX t = e tx x 1 e x dx = x 1 e dx.

0

0

1 t

, dx = dy

1 t

, and y [0, ), the above

expression is equal to

1 1 1

( ) 0

y 1 e y dy = ,

(1 t ) (1 t )

d

() d

( ) ( )

MX t = 1 t = = E X ,

dt t =0

dt t =0

156 6 Characteristic Functions, Moment Generating Functions and Related Theorems

and

d2

() d

( ) ( ) ( )

1 2

MX t = 1 t = + 1 2 1 t

dt 2 t =0

dt t =0 t =0

( ) (

= + 1 = EX , so that X = .

2 2

) 2

( ) 2

In particular, for = 2r and = 2, we get the m.g.f. of the 2r, and its mean and

variance; namely,

() ( 1

) ( ) ( )

r 2

MX t = 1 2t , E X = r , 2 X = 2r .

, t<

2

For = 1 and = 1 , we obtain the m.g.f. of the Negative Exponential

distribution, and its mean and variance; namely

MX t =() t

1

, t < , EX = , 2 X = 2 .

1

( )

5. Let X have the Cauchy distribution with parameters and , and

without loss of generality, let = 0, = 1. Then the MX(t) exists only for t = 0.

In fact.

() ( )

1 1

M X t = E e tX = e tx dx

1 + x2

1

> e tx

0

1

1+ x 2

1

dx > tx

0

1

1 + x2

dx ( )

if t > 0, since ez > z, for z > 0, and this equals

t

2

2 x dx

1+ x

20

=

t du

2 1 u

=

t

lim log u .

2 x ( )

Thus for t > 0, MX(t) obviously is equal to . If t < 0, by using the limits , 0

in the integral, we again reach the conclusion that MX(t) = (see Exercise

6.5.9).

REMARK 4 The examples just discussed exhibit all three cases regarding the

existence or nonexistence of an m.g.f. In Examples 1 and 3, the m.g.f.s exist

for all t !; in Examples 2 and 4, the m.g.f.s exist for proper subsets of !; and

in Example 5, the m.g.f. exists only for t = 0.

For an r.v. X, we also define what is known as its factorial moment

generating function. More precisely, the factorial m.g.f. X (or just when no

confusion is possible) of an r.v. X is defined by:

() ( )

X t = E t X , t ! , if E t X exists. ( )

This function is sometimes referred to as the Mellin or MellinStieltjes trans-

form of f. Clearly, X(t) = MX(log t) for t > 0.

Formally, the nth factorial moment of an r.v. X is taken from its factorial

m.g.f. by differentiation as follows:

6.5 The Moment Generating Function 157

dn

dt n

X t ()

t =1

[ ( ) (

= E X X 1 X n+1 . )]

In fact,

n X

dn

dt n

()

X t =

dn

dt n

E t X

=( )

E n t = E X X 1 X n+1 t

t

[ (

X n

, ) ( ) ]

provided Lemma D applies, so that the interchange of the order of differen-

tiation and expectation is valid. Hence

dn

dt n

X t ()

t =1

[ ( ) (

= E X X 1 X n+1 . )] (9)

REMARK 5 The factorial m.g.f. derives its name from the property just estab-

lished. As has already been seen in the first two examples in Section 2 of

Chapter 5, factorial moments are especially valuable in calculating the vari-

ance of discrete r.v.s. Indeed, since

( ) ( ) ( ) 2

2 X = E X 2 EX , and E X 2 = E X X 1 + E X , ( ) [ ( )] ( )

we get

( ) [ (

2 X = E X X 1 + E X EX ; )] ( ) ( )

2

m.g.f. up to order two.

Below we derive some factorial m.g.f.s. Property (9) (for n = 2) is valid in

all these examples, although no specific justification will be provided.

1. If X B(n, p), then X(t) = (pt + q)n, t !. In fact,

n

n n

n

() ( ) ( )

x n

X t = t x p x q n x = pt q n x = pt + q .

x= 0 x x= 0 x

Then

d2

() ( ) ( ) ( )

n 2

X t = n n 1 p 2 pt + q = n n 1 p2 ,

dt 2 t =1

2. If X P(), then X(t) = et , t !. In fact,

( )

x

t

x

()

X t = t e

x =0

x

x!

= e

x = 0 x!

= e e t = e t , t ! .

158 6 Characteristic Functions, Moment Generating Functions and Related Theorems

Hence

d2

dt 2

X t () = 2 e t

t =1

= 2 , so that 2 X = 2 + 2 = . ( )

t =1

The m.g.f. of an r. vector X or the joint m.g.f. of the r.v.s X1, . . . , Xk,

denoted by MX or MX1, . . . , Xk, is defined by:

1 k

( ) (

M X , , X t1 , , t k = E e t X + t X , t j ! , j = 1, 2, , k,

1 1 k k

)

for those tjs in ! for which this expectation exists. If MX1, . . . , Xk(t1, . . . , tk) exists,

then formally X1, . . . , Xk(t1, . . . , tk) = MX1, . . . , Xk(it1, . . . , itk) and properties analo-

gous to (i)(vii), (viii) in Theorem 1 hold true under suitable conditions. In

particular,

n + +n

(t , , t ) ( )

1 k

MX , , Xk 1 k = E X 1n 1 X knk , (10)

t1n 1 t knk 1

t 1 = = tk = 0

Below, we present two examples of m.g.f.s of r. vectors.

1. If the r.v.s X1, . . . , Xk have jointly the Multinomial distribution with

parameters n and p1, . . . , pk, then

(t , , t ) = ( p e ),

n

t1

MX , ,

1 Xk 1 k 1 + + pk e t k

t j ! , j = 1, , k.

In fact,

MX1 , , Xk (t , , t ) = Ee

1 k

t 1 X 1 + +tk X k

= e 1

t X 1 + +tk X k n!

x1! xk !

p1x1 pkxk

n!

( ) ( )

x1 xk

= p1e t1 pk e tk

x1! xk !

( ),

n

= p1e t1 + + pk e tk

Clearly, the above derivations hold true for all tj !, j = 1, . . . , k.

2. If the r.v.s X1 and X2 have the Bivariate Normal distribution with

parameters 1, 2, 21, 22 and , then their joint m.g.f. is

MX 1 ,X 2

(t , t ) = exp t

1 2

1 1 + 2t 2 +

2

(

1 2 2

1 t1 + 2 1 2 t1t 2 + 22 t 22 ), t , t

1 2 ! . (11)

the matrix approach, which is more elegant and compact. Recall that the joint

p.d.f. of X1 and X2 is given by

6.5 The Moment Generating Function 159

(

f x1 , x 2 )

2

1 x1 1 x1 1 x 2 2 x 2 2

2

1

= exp 2 + .

2 1 2 1 2 2 1 1 2 2

Set x = (x1 x2), = (1 2), and

12

12 2 .

=

1 2 2

|, is |

1 2 1 2

2

1 = .

1 2 12

Therefore

22 1 2 x1 1

( )

x 1 x = ( ) 1

(

x1 1 x 2 2

1 2

)

12 x 2 2

1

(

2 x

) ( )( ) ( )

2 1 2 x1 1 x 2 2 + 12 x 2 2

2 2

=

1

1

2 2

2 ( 2

)

2 1 1

1 x 2 x1 1 x 2 2 x 2 2

2

= 1 1

2 +

.

1 2 1 1 2 2

1

()

f x =

1

( )

exp x 1 x ( ) .

2

1 2

2

In this form, is the mean vector of X = (X1 X2), and is the covariance matrix

of X.

Next, for t = (t1 t2), we have

()

M X t = Ee t X = 2 exp t x f x dx

!

( )()

! exp t x 2 ( x ) ( x )dx.

1 1 1

= 1 2 2

(11)

2

160 6 Characteristic Functions, Moment Generating Functions and Related Theorems

1

2

1

2

[ ( 1

)

t + t t 2 t + t t 2 t x + x x . ( )] (13)

Focus on the quantity in the bracket, carry out the multiplication, and observe

1) = 1, xt = tx, t = t, and x 1 =

that = , ( 1x, to obtain

( ) ( ) [ ( ) ( (

2 t + t t 2t x + x 1 x = x + t 1 x + t . ))] (14)

MX t()

1

= exp t + t t

2

1

2 1 2 ! 2

1

( ( )) ( (

exp x + t 1 x + t dx.

2

))

However, the second factor above is equal to 1, since it is the integral of a

Bivariate Normal distribution with mean vector + t and covariance matrix

. Thus

()

1

M X t = exp t + t t .

2

(15)

Observing that

12 1 2 t1

( )

t t = t1 t 2 = 1 t1 + 2 1 2 t1t 2 + 2 t 2 ,

1 2 2 t 2

2

2 2 2 2

Exercises

6.5.1 Derive the m.g.f. of the r.v. X which denotes the number of spots that

turn up when a balanced die is rolled.

Derive its m.g.f. and factorial m.g.f., M(t) and (t), respectively, for those ts

for which they exist. Then calculate EX, E[X(X 1)] and 2(X), provided they

are finite.

Derive its m.g.f. and factorial m.g.f., M(t) and (t), respectively, for those ts

for which they exist. Then calculate EX, E[X(X 1)] and 2(X), provided they

are finite.

6.5 The Moment GeneratingExercises

Function 161

6.5.4 Let X be an r.v. with p.d.f. f given by f(x) = e(x )I(,)(x). Find its

m.g.f. M(t) for those ts for which it exists. Then calculate EX and 2(X),

provided they are finite.

6.5.5 Let X be an r.v. distributed as B(n, p). Use its factorial m.g.f. in order

to calculate its kth factorial moment. Compare with Exercise 5.2.1 in Chapter

5.

6.5.6 Let X be an r.v. distributed as P(). Use its factorial m.g.f. in order to

calculate its kth factorial moment. Compare with Exercise 5.2.4 in Chapter 5.

6.5.7 Let X be an r.v. distributed as Negative Binomial with parameters

r and p.

i) Show that its m.g.f and factorial m.g.f., M(t) and (t), respectively, are

given by

pr pr

()

MX t = , t < log q, X t = () , t <

1

;

(1 qe ) (1 qt ) q

r r

t

iii) Find the quantities mentioned in parts (i) and (ii) for the Geometric

distribution.

6.5.8 Let X be an r.v. distributed as U(, ).

ii) Show that its m.g.f., M, is given by

e t e t

Mt = () ;

(

t )

ii) By differentiation, show that EX = +2 and 2(X) = ( ) .

2

12

6.5.9 Refer to Example 3 in the Continuous case and show that MX(t) = for

t < 0 as asserted there.

2

6.5.10 Let X be an r.v. with m.g.f. M given by M(t) = e t + t , t ! ( !,

> 0). Find the ch.f. of X and identify its p.d.f. Also use the ch.f. of X in order

to calculate EX4.

6.5.11 For an r.v. X, define the function by (t) = E(1 + t)X for those ts for

which E(1 + t)X is finite. Then, if the nth factorial moment of X is finite, show

that

(d n

) ()

dt n t

t =0

[ ( )

= E X X 1 X n+1 .( )]

6.5.12 Refer to the previous exercise and let X be P(). Derive (t) and use

it in order to show that the nth factorial moment of X is n.

162 6 Characteristic Functions, Moment Generating Functions and Related Theorems

6.5.13 Let X be an r.v. with m.g.f. M and set K(t) = log M(t) for those ts for

which M(t) exists. Furthermore, suppose that EX = and 2(X) = 2 are both

finite. Then show that

d2

d

dt

Kt() = and

dt 2

Kt () = 2.

t =0 t =0

(The function K just defined is called the cumulant generating function of X.)

6.5.14 Let X be an r.v. such that EX n is finite for all n = 1, 2, . . . . Use the

expansion

xn

ex =

n = 0 n!

in order to show that, under appropriate conditions, one has that the m.g.f. of

X is given by

( ) tn! .

n

()

M t = EX n

n =0

6.5.15 If X is an r.v. such that EX n = n!, then use the previous exercise in

order to find the m.g.f. M(t) of X for those ts for which it exists. Also find the

ch.f. of X and from this, deduce the distribution of X.

6.5.16 Let X be an r.v. such that

EX 2k =

(2k)! , EX 2k +1 = 0,

2 k k!

k = 0, 1, . . . . Find the m.g.f. of X and also its ch.f. Then deduce the distribution

of X. (Use Exercise 6.5.14)

6.5.17 Let X1, X2 be two r.v.s with m.f.g. given by

2

( )

1

3

( 1

) (

)

M t1 , t 2 = e t1 + t2 + 1 + e t1 + e t2 , t1 , t 2 ! .

6

Calculate EX1, 2(X1) and Cov(X1, X2), provided they are finite.

6.5.18 Refer to Exercise 4.2.5. in Chapter 4 and find the joint m.g.f.

M(t1, t2, t3) of the r.v.s X1, X2, X3 for those t1, t2, t3 for which it exists. Also find

their joint ch.f. and use it in order to calculate E(X1X2X3), provided the

assumptions of Theorem 1 (vii) are met.

6.5.19 Refer to the previous exercise and derive the m.g.f. M(t) of the r.v.

g(X1, X2, X3) = X1 + X2 + X3 for those ts for which it exists. From this, deduce

the distribution of g.

6.5 The Moment GeneratingExercises

Function 163

6.5.20 Let X1, X2 be two r.v.s with m.g.f. M and set K(t1, t2) = log M(t1, t2) for

those t1, t2 for which M(t1, t2) exists. Furthermore, suppose that expectations,

variances, and covariances of these r.v.s are all finite. Then show that for

j = 1, 2,

2

t j

(

K t1 , t 2 ) = EX j ,

t j2

(

K t1 , t 2 ) ( )

= 2 Xj ,

t 1 =t2 = 0 t 1 =t2 = 0

2

t1t 2

K t1 , t 2 ( ) (

= Cov X 1 , X 2 . )

t 1 =t2 = 0

6.5.21 Suppose the r.v.s X1, . . . , Xk have the Multinomial distribution with

parameters n and p1, . . . , pk, and let i, j, be arbitrary but fixed, 1 i < j k.

Consider the r.v.s Xi, Xj, and set X = n Xi Xj, so that these r.v.s have the

Multinomial distribution with parameters n and pi, pj, p, where p = 1 pi pj.

ii) Write out the joint m.g.f. of Xi, Xj, X, and by differentiation, determine the

E(XiXj);

ii) Calculate the covariance of Xi, Xj, Cov(Xi, Xj), and show that it is negative.

6.5.22 If the r.v.s X1 and X2 have the Bivariate Normal distribution with

parameters 1, 2, 21, 22 and , show that Cov(X1, X2) 0 if 0, and

Cov(X1, X2) < 0 if < 0. Note: Two r.v.s X1, X2 for which Fx ,x (X1, X2) 1 2

Fx (X1)Fx (X2) 0, for all X1, X2 in !, or Fx ,x (X1, X2) Fx (X1)Fx (X2) 0, for

1 2 1 2 1 2

all X1, X2 in !, are said to be positively quadrant dependent or negatively

quadrant dependent, respectively. In particular, if X1 and X2 have the Bivariate

Normal distribution, it can be seen that they are positively quadrant depend-

ent or negatively quadrant dependent according to whether 0 or < 0.

6.5.23 Verify the validity of relation (13).

6.5.24

ii) If the r.v.s X1 and X2 have the Bivariate Normal distribution with param-

eters 1, 2, 21, 22 and , use their joint m.g.f. given by (11) and property

(10) in order to determine E(X1X2);

ii) Show that is, indeed, the correlation coefficient of X1 and X2.

6.5.25 Both parts of Exercise 6.4.1 hold true if the ch.f.s involved are re-

placed by m.g.f.s, provided, of course, that these m.g.f.s exist.

ii) Use Exercise 6.4.1 for k = 2 and formulated in terms of m.g.f.s in order to

show that the r.v.s X1 and X2 have a Bivariate Normal distribution if and

only if for every c1, c2 !, Yc = c1X1 + c2X2 is normally distributed;

ii) In either case, show that c1X1 + c2X2 + c3 is also normally distributed for any

c3 !.

164 7 Stochastic Independence with Some Applications

Chapter 7

Applications

Let S be a sample space, consider a class of events associated with this space,

and let P be a probability function defined on the class of events. In Chapter

2 (Section 2.3), the concept of independence of events was defined and was

heavily used there, as well as in subsequent chapters. Independence carries

over to r.v.s also, and is the most basic assumption made in this book. Inde-

pendence of r.v.s, in essence, reduces to that of events, as will be seen below.

In this section, the not-so-rigorous definition of independence of r.v.s is pre-

sented, and two criteria of independence are also discussed. A third criterion

of independence, and several applications, based primarily on independence,

are discussed in subsequent sections. A rigorous treatment of some results is

presented in Section 7.4.

DEFINITION 1 The r.v.s Xj, j = 1, . . . , k are said to be independent if, for sets Bj !, j = 1, . . . ,

k, it holds

k

( ) (

P X j Bj , j = 1, , k = P X j Bj .

j =1

)

The r.v.s Xj, j = 1, 2, . . . are said to be independent if every finite subcollection

of them is a collection of independent r.v.s. Non-independent r.v.s are said to

be dependent. (See also Definition 3 in Section 7.4, and the comment following

it.)

REMARK 1 (i) The sets Bj, j = 1, . . . , k may not be chosen entirely arbitrar-

ily, but there is plenty of leeway in their choice. For example, taking Bj = (,

xj], xj !, j = 1, . . . , k would be sufficient. (See Lemma 3 in Section 7.4.)

(ii) Definition 1 (as well as Definition 3 in Section 7.4) also applies to m-

dimensional r. vectors when ! (and B in Definition 3) is replaced by ! m (Bm).

164

7.1 Stochastic Independence: Criteria of Independence 165

THEOREM 1 (Factorization Theorem) The r.v.s Xj, j = 1, . . . , k are independent if and only

if any one of the following two (equivalent) conditions holds:

k

1 k

( )

i) FX , , X x1 , , xk = FX x j , for all

j =1

j

( ) x j ! , j = 1, , k.

1 k

( )

ii) f X , , X x1 , , xk = f X x j , for all

j =1

j

( ) x j ! , j = 1, , k.

PROOF

ii) If Xj, j = 1, , k are independent, then

k

( )

P X j Bj , j = 1, , k = P X j Bj , Bj ! , j = 1, , k.

j =1

( )

In particular, this is true for Bj = (, xj], xj !, j = 1, . . . , k which gives

k

1 k

(

FX , , X x1 , , xk = FX x j . ) j =1

j

( )

The proof of the converse is a deep probability result, and will, of course,

be omitted. Some relevant comments will be made in Section 7.4, Lemma 3.

ii) For the discrete case, we set Bj = {xj}, where xj is in the range of Xj, j = 1, . . . ,

k. Then if Xj, j = 1, . . . , k are independent, we get

k

(

P X 1 = x1 , , X k = xk = P X j = x j , ) j =1

( )

or

k

1 k

(

f X , , X x1 , , xk = f X x j . ) j =1

j

( )

Let now

k

1 k

(

f X , , X x1 , , xk = f X x j . ) j =1

j

( )

Then for any sets Bj = (, yj], yj !, j = 1, . . . , k, we get

B1

B

1 k

(

f X , , X x1 , , xk = ) B1

B

1

( )

f X x1 f X xk k

( )

k k

k

= f X x j ,

j =1

j

( )

B j

or

k

1 k

(

FX , , X y1 , , yk = FX y j . ) j =1

j

( )

Therefore Xj, j = 1, . . . , k are independent by (i). For the continuous case,

we have: Let

k

1 k

(

f X , , X x1 , , xk = f X x j ) j =1

j

( )

and let

166 7 Stochastic Independence with Some Applications

( ]

C j = , y j , y j ! . j = 1, , k.

Then integrating both sides of this last relationship over the set C1

Ck, we get

k

1 k

(

FX , , X y1 , , yk = FX y j , ) j =1

j

( )

so that Xj, j = 1, . . . , k are independent by (i). Next, assume that

k

1 k

(

FX , , X x1 , , xk = FX x j ) j =1

j

( )

(that is, the Xjs are independent). Then differentiating both sides, we get

k

1 k

( )

f X , , X x1 , , xk = f X x j .

j =1

j

( )

REMARK 2 It is noted that this step also is justifiable (by means of calculus)

for the continuity points of the p.d.f. only.

Consider independent r.v.s and suppose that gj is a function of the jth r.v.

alone. Then it seems intuitively clear that the r.v.s gj(Xj), j = 1, . . . , k ought to

be independent. This is, actually, true and is the content of the following

LEMMA 1 For j = 1, . . . , k, let the r.v.s Xj be independent and consider (measurable)

functions gj : ! !, so that gj(Xj), j = 1, . . . , k are r.v.s. Then the r.v.s gj(Xj),

j = 1, . . . , k are also independent. The same conclusion holds if the r.v.s are

replaced by m-dimensional r. vectors, and the functions gj, j = 1, . . . , k are

defined on ! m into !. (That is, functions of independent r.v.s (r. vectors)

are independent r.v.s.)

PROOF See Section 7.4.

Independence of r.v.s also has the following consequence stated as a

lemma. Both this lemma, as well as Lemma 1, are needed in the proof of

Theorem 1 below.

LEMMA 2 Consider the r.v.s Xj, j = 1, . . . , k and let gj : ! ! be (measurable) functions,

so that gj(Xj), j = 1, . . . , k are r.v.s. Then, if the r.v.s Xj, j = 1, . . . , k are

independent, we have

k k

( )

E g j X j = E g j X j ,

j = 1 j = 1

[ ( )]

provided the expectations considered exist. The same conclusion holds if the

gjs are complex-valued.

PROOF See Section 7.2.

REMARK 3 The converse of the above statement need not be true as will be

seen later by examples.

THEOREM 1 (Factorization Theorem) The r.v.s Xj, j = 1, . . . , k are independent if and only if:

k

1 k

( ) j =1

j

( )

X , , X t1 , , t k = X t j , for all t j ! , j = 1, , k.

7.1 Stochastic Independence: Criteria of Independence 167

k

1 k

(

f X , , X x1 , , xk = f X x j . ) j =1

j

( )

Hence

k k it X k

(

)

X , , X x1 , , xk = E exp i t j X j = E e = Ee

it X j j j j

1 k

j =1 j =1 j =1

by Lemmas 1 and 2, and this is kj=1X (tj). Let us assume now that j

k

1 k

(

X , , X t1 , , t k = X t j . ) j =1

j

( )

For the discrete case, we have (see Theorem 2(i) in Chapter 6)

f X x j = lim j

( )

1 T it x

T 2T T

e X t j dt j , j = 1, , k, j j

j

( )

and for the multidimensional case, we have (see Theorem 2(i) in Chapter 6)

1

k

k

f X 1 , , X k ( x1 , , x k ) = lim

T T

exp i t j x j

T 2T T T

j =1

X 1 , , X k (t1 , , t k )dt1 dt k

1

k

k k

exp i t j x j X j t j ( dt1 dt k ) ( )

T T

= lim

T 2T

T T

j =1 j =1

1 T it j x j k

( )

k

= lim e X j t j dt j = f X x .

j =1

T 2T T

j =1 j ( j )

That is, Xj, j = 1, . . . , k are independent by Theorem 1(ii). For the continuous

case, we have

it j h

1 T 1 e

( )

h 0 T 2 T

f X x j = lim lim

j

it j h

it h

e X t j dt j , j = 1, , k, j

j

( )

and for the multidimensional case, we have (see Theorem 2(ii) in Chapter 6)

k

1 k 1 e it j h

( )

T T

T

it j x j

f X 1 , , X k x1 , , xk = lim lim e

h 0 T 2 T it j h

j =1

(

X 1 , , X k t1 , , t k dt1 dt k )

k

1 1 e it j h it j x j

k

( )

T T

= lim lim

h 0 T 2 T e Xj t j

it j h

T

j =1

dt1 dt k

k it h

1 e j it j x j

1

( )

T

= lim lim T it j h e X j t j dt j

h 0 T 2

j =1

k

= fX j ( xj ) ,

j =1

168 7 Stochastic Independence with Some Applications

1(ii).

REMARK 4 A version of this theorem involving m.g.f.s can be formulated, if

the m.g.f.s exist.

COROLLARY Let X1, X2 have the Bivariate Normal distribution. Then X1, X2 are indepen-

dent if and only if they are uncorrelated.

PROOF We have seen that (see Bivariate Normal in Section 3.3 of Chapter 3)

fX 1, X2 (x , x ) =

1 2

1

e q 2 ,

2 1 2 1 2

where

1 x 2 x1 2 x2 2 x2 2

2

q= 1 1

2

,

1 2 1 1 2 2

and

x 2

( ) x

( )

2

( )

fX 1 x1 =

1

exp

1

2 2

1 , f x =

X2 2

1

exp 2

2

( )

2

2 .

2 1 1

2 2 2

Thus, if X1, X2 are uncorrelated, so that = 0, then

( ) ( )

fX 1 , X 2 x1 , x 2 = fX 1 x1 fX 2 x 2 , ( )

that is, X1, X2 are independent. The converse is always true by Corollary 1 in

Section 7.2.

Exercises

7.1.1 Let Xj, j = 1, . . . , n be i.i.d. r.v.s with p.d.f. f and d.f. F. Set

(

X (1) = min X 1 , , X n , ) (

X n = max X 1 , , X n ; )

that is,

() [ ()

X (1) s = min X 1 s , , X n s , ( )] () [ () ( )]

X ( n ) s = max X 1 s , , X n s .

Then express the d.f. and p.d.f. of X(1), X(n) in terms of f and F.

7.1.2 Let the r.v.s X1, X2 have p.d.f. f given by f(x1, x2) = I(0,1) (0,1)(x1, x2).

ii) Show that X1, X2 are independent and identify their common distribution;

ii) Find the following probabilities: P(X1 + X2 < 13 ), P( X12 + X 22 < 14 ),

P(X1X2 > 12 ).

7.1.3 Let X1, X2 be two r.v.s with p.d.f. f given by f(x1, x2) = g(x1)h(x2).

Exercises

7.1 Stochastic Independence: Criteria of Independence 169

ii) Derive the p.d.f. of X1 and X2 and show that X1, X2 are independent;

ii) Calculate the probability P(X1 > X2) if g = h and h is of the continuous

type.

7.1.4 Let X1, X2, X3 be r.v.s with p.d.f. f given by f(x1, x2, x3) = 8x1x2x3 IA(x1,

x2, x3), where A = (0, 1) (0, 1) (0, 1).

ii) Show that these r.v.s are independent;

ii) Calculate the probability P(X1 < X2 < X3).

7.1.5 Let X1, X2 be two r.v.s with p.d.f f given by f(x1, x2) = cIA(x1, x2), where

A = {(x1, x2) !2; x12 + x 22 9}.

ii) Determine the constant c;

ii) Show that X1, X2 are dependent.

7.1.6 Let the r.v.s X1, X2, X3 be jointly distributed with p.d.f. f given by

( )1

(

f x1 , x 2 , x3 = I A x1 , x 2 , x3 ,

4

)

where

{( )( )( )(

A = 1, 0, 0 , 0, 1, 0 , 0, 0, 1 , 1, 1, 1 . )}

Then show that

ii) Xi, Xj, i j, are independent;

ii) X1, X2, X3 are dependent.

7.1.7 Refer to Exercise 4.2.5 in Chapter 4 and show that the r.v.s X1, X2, X3

are independent. Utilize this result in order to find the p.d.f. of X1 + X2 and X1

+ X2 + X3 .

7.1.8 Let Xj, j = 1, . . . , n be i.i.d. r.v.s with p.d.f. f and let B be a (Borel) set

in !.

iii) In terms of f, express the probability that at least k of the Xs lie in B for

some fixed k with 1 k n;

iii) Simplify this expression if f is the Negative Exponential p.d.f. with param-

eter and B = (1/, );

iii) Find a numerical answer for n = 10, k = 5, = 12 .

7.1.9 Let X1, X2 be two independent r.v.s and let g: ! ! be measurable.

Let also Eg(X2) be finite. Then show that E[g(X2) | X1 = x1] = Eg(X2).

7.1.10 If Xj, j = 1, . . . , n are i.i.d. r.v.s with ch.f. and sample mean X ,

express the ch.f. of X in terms of .

7.1.11 For two i.i.d. r.v.s X1, X2, show that X X (t) = |X (t)|2, t !. (Hint:

1 2 1

170 7 Stochastic Independence with Some Applications

7.1.12 Let X1, X2 be two r.v.s with joint and marginal ch.f.s X ,X , X and X . 1 1 1 2

X 1 ,X2

(t , t ) = (t ) (t ),

1 2 X1 1 X2 2 t1 , t 2 !.

X 1 ,X2

(t, t ) = (t ) (t ), X1 X2 t ! ,

We now proceed with the proof of Lemma 2.

PROOF OF LEMMA 2 Suppose that the r.v.s involved are continuous, so that

we use integrals. Replace integrals by sums in the discrete case. Thus,

k

( ) ( ) ( )

E g j X j = g1 x1 gk xk

j =1

) (

f X , , X x1 , , xk dx1 dxk

1 k

= g ( x ) g ( x ) f ( x ) f ( x )dx dx

1 1 k k X1 1 Xk k 1 k

(by independence)

= g ( x ) f ( x )dx g ( x ) f ( x )dx

1 1 X1 1 1 k k Xk k k

= E[ g ( X )] E[ g ( X )].

1 1 k k

Now suppose that the gjs are complex-valued, and for simplicity, set gj(Xj) = Yj

= Yj1 + Yj2, j = 1, . . . , k. For k = 2,

( ) [(

E Y1Y2 = E Y11 + iY12 Y21 + iY22 )( )]

( ) (

= E Y11Y21 Y12Y22 + iE Y11Y22 + Y12Y21)

= [ E(Y Y ) E(Y Y )] + i[ E(Y Y ) + E(Y Y )]

11 21 12 22 11 22 12 21

= [( EY )( EY ) ( EY )( EY )] + i[( EY )( EY ) ( EY )( EY )]

11 21 12 22 11 22 12 21

= ( EY + iEY )( EY + iEY ) = ( EY )( EY ).

11 12 21 22 1 2

Indeed,

7.2 Independence:

7.1 Stochastic Proof of Lemma 2 andof

Criteria Related Results

Independence 171

( ) [(

E Y1 Ym+1 = E Y1 Ym Ym+1 ) ]

= E(Y Y )( EY ) (by the part just established)

1 m m+1

1 m m+1

COROLLARY 1 The covariance of an r.v. X and of any other r.v. which is equal to a constant

c (with probability 1) is equal to 0; that is, Cov(X, c) = 0.

PROOF Cov(X, c) = E(cX) (Ec)(EX) = cEX cEX = 0.

COROLLARY 2 If the r.v.s X1 and X2 are independent, then they have covariance equal to 0,

provided their second moments are finite. In particular, if their variances are

also positive, then they are uncorrelated.

PROOF In fact,

( ) (

Cov X1 , X 2 = E X1 X 2 EX1 EX 2 ) ( )( )

= ( EX )( EX ) ( EX )( EX ) = 0,

1 2 1 2

REMARK 5 The converse of the above corollary need not be true. Thus

uncorrelated r.v.s in general are not independent. (See, however, the corol-

lary to Theorem 1 after the proof of part (iii).)

COROLLARY 3 i) For any k r.v.s Xj, j = 1, . . . , k with finite second moments and variances

2j = 2(Xj), and any constants cj, j = 1, . . . , k, it holds:

k k

2 c j X j = c 2j j2 + ci c j Cov X i , X j

j =1 j =1 1i j k

( )

k

= c 2j j2 + 2

j =1

1i < j k

(

ci c j Cov X i , X j . )

ii) If also j > 0, j = 1, . . . , k, and ij = (Xi, Xj), i j, then:

k k

2 c j X j = c 2j j2 + ci c j i j ij

j =1 j =1 1i j k

k

= c 2j j2 + 2 ci c j i j ij .

j =1 1i < j k

(are pairwise uncorrelated), then:

iii) 2(kj= 1cjXj) = kj= 1c2j 2j, and

iii) 2(kj= 1Xj) = kj= 1 2j (Bienaym equality).

172 7 Stochastic Independence with Some Applications

PROOF

iii) Indeed,

2

k k k

c j X j = E c j X j E c j X j

2

j =1 j = 1 j =1

2

k

j = 1

(

= E c j X j EX j )

k

( ) + c c (X )( )

2

= E c 2j X j EX j i j i EX i X j EX j

j = 1 i j

k

= c 2j j2 +

j =1

1i j k

(

ci c j Cov X i , X j )

k

= c 2j j2 + 2

j =1

1i < j k

(

ci c j Cov X i , X j )

(since Cov( X , X ) = Cov( X , X )).

i j j i

This establishes part (i). Part (ii) follows by the fact that Cov(Xi, Xj) =

ijij = jiji.

iii) Here Cov (Xi, Xj) = 0, i j, either because of independence and Corollary

2, or ij = 0, in case j > 0, j = 1, . . . , k. Then the assertion follows from

either part (i) or part (ii), respectively.

iii) Follows from part (iii) for c1 = = ck = 1.

Exercises

7.2.1 For any k r.v.s Xj, j = 1, . . . , k for which E(Xj) = (finite) j = 1, . . . , k,

show that

k k

(X ) = (X ) ( ) ( )

2 2 2 2

j j X +k X = kS 2 + k X ,

j =1 j =1

where

2

1 k 1 k

j

X=

k j =1

X and S 2

= Xj X .

k j =1

( )

## Molto più che documenti.

Scopri tutto ciò che Scribd ha da offrire, inclusi libri e audiolibri dei maggiori editori.

Annulla in qualsiasi momento.